---
title: "Calculations across the site: pre-registered specification"
url: https://tipsheet.markets/research/site_calculations_spec/
description: "Pre-registered research specification and results log: Calculations across the site: pre-registered specification."
updated: 2026-10-29
publisher: tipsheet (https://tipsheet.markets/)
cite: "tipsheet, “Calculations across the site: pre-registered specification”, 29 October 2026, https://tipsheet.markets/research/site_calculations_spec/"
---

[Pre-registered research](https://tipsheet.markets/research/)

# Calculations across the site: pre-registered specification

Written 2026-10-02 by lane 2, **before any of these was computed**. Each block states the question, the data, the method, the overlap handling and a prediction to be graded. Changes go under “Amendments” with the date and the reason.

Common rules:

- **No look-ahead:** every signal uses only data published before the date it is used (one-month lag for monthly macro and valuation series unless stated).
- **Overlap:** forward returns over h months computed every month overlap by h−1 months. I report the effective number of independent periods (sample months ÷ h), Newey-West (Hansen-Hodrick style) standard errors with h lags, and the same statistic on non-overlapping subsamples (every h-th month, for each of the h possible offsets, with the range across offsets).
- **Nulls are results.** A calculation that finds nothing is published with the same prominence as one that finds something.

## C1. Valuation and subsequent returns (CAPE, earnings yield, yield gap)

- **Question:** did starting valuation tell you anything about the next 5 and 10 years of Indian equity returns?
- **Data:** IIMA Sensex CAPE10 (month-end, from April 2000) and Nifty 500 CAPE10 (from April 2006) from `docs/reference/iima_india_cape_*.csv`; Nifty 500 trailing earnings yield and the yield gap (earnings yield minus 10-year G-sec) from `valuation_nifty_500.parquet` (1999 on); forward returns from the Nifty 500 TRI, nominal and real (all-India CPI, spliced).
- **Method:** regress the annualised forward 5-year and 10-year real return on log CAPE (and separately on the earnings yield and the yield gap) using the value known at the start month (one-month lag). Report the slope, R², Newey-West t-statistic with 60 or 120 lags, the non-overlapping-offset range of slopes, and a table of median forward returns by CAPE tercile with the number of independent periods in each.
- **Prediction C1:** the slope on log CAPE is negative at both horizons, but at 10 years the sample holds fewer than three independent periods, so I will not call it significant whatever the t-statistic says.

## C2. Breadth thrusts and their record

- **Question:** do NSE breadth thrusts (Zweig 1986) precede above-normal returns?
- **Data:** daily advances and declines of NSE stocks from bhavcopy (`breadth_stocks_daily.parquet`), Nifty 500 TRI.
- **Method:** the 10-session EMA of advances / (advances + declines) rises from below 0.40 to above 0.615 within 10 sessions; a new signal needs 60 sessions since the last. For each signal, forward 1, 3, 6 and 12-month Nifty 500 returns, against the unconditional distribution of forward returns over the same sample. Significance by a block bootstrap of signal dates (random dates with the same count, mean block 63 sessions).
- **Prediction C2:** fewer than 15 signals since 1995; the average forward 6-month return after a signal is above the unconditional average, but the bootstrap p-value is above 0.05.

## C3. Flows and returns

- **Question:** do foreign (FPI) and domestic institutional (DII) flows lead market returns, or follow them?
- **Data:** daily net flows (`flows_daily.parquet`, NSDL FPI and exchange-reported DII), weekly sums scaled by Nifty 500 market capitalisation where available; Nifty 500 TRI weekly returns.
- **Method:** cross-correlations of weekly flows and returns at leads and lags of up to 8 weeks, and bivariate Granger tests (lag length by BIC, up to 4) in each direction, on the full sample and two halves. Caveats stated: Granger causality is predictability, not cause; flows and returns are measured with timing mismatches.
- **Prediction C3:** returns predict next week’s FPI flows (positive, significant) more strongly than flows predict next week’s returns; DII flows respond negatively to past returns (domestic institutions buy dips).

## C4. The IPO cycle and market returns

- **Question:** does a hot IPO market precede weak returns (Baker and Wurgler 2000, the equity share in new issues)?
- **Data:** monthly IPO count and money raised (`ipo_fundraising_monthly.parquet`), Nifty 500 TRI and its market capitalisation proxy.
- **Method:** IPO money raised over the trailing 12 months as a share of market capitalisation, against forward 12 and 36-month Nifty 500 returns; overlap handling as above.
- **Prediction C4:** a negative relation at 36 months, not significant after accounting for overlap.

## C5. Leverage (MTF) and positioning against drawdowns

- **Question:** does a rising margin-trading book or crowded futures positioning precede drawdowns?
- **Data:** the MTF book (2004 on) and participant open interest from `leverage` (lane 1 tables).
- **Method:** 3-month growth of the MTF book (and its level relative to market capitalisation) against the probability of a 10% Nifty 500 drawdown in the next 3 months (logistic regression) and forward 3-month returns; overlap handling as above.
- **Prediction C5:** high MTF growth is associated with a higher drawdown probability, with a confidence interval that includes no effect.

## C6. SIP outcome distribution

Covered by `sip_studies_spec.md`.

## C7. A survivorship-free factor library from bhavcopy

- **Question:** what did momentum, low volatility, short-term reversal and size earn in Indian stocks since 1995, net of realistic costs, when delisted stocks are kept?
- **Data:** split- and bonus-adjusted NSE bhavcopy prices with delistings kept (`equity_prices_adjusted.parquet`). Value and quality need fundamentals the warehouse does not hold yet; size needs shares outstanding (request to lane 1).
- **Method:** monthly rebalanced long-only top-quintile and long-short quintile portfolios within the top 500 stocks by traded value (formation at month-end, trade at the next close): momentum 12-1 (Jegadeesh and Titman 1993), 252-day volatility (lowest quintile), 1-month reversal; equal and liquidity-capped weights; costs 0.3% each way for the top 200 and 0.6% beyond, plus impact proportional to trade size over traded value. Delisted stocks: last traded price, then a −30% delisting return for compulsory delistings where the reason is known, 0% otherwise, stated.
- **Prediction C7:** momentum’s long-short premium is positive before costs and roughly halves after costs; low volatility has a higher Sharpe ratio than the market with a lower return; short-term reversal does not survive costs.

## C8. Fear & Greed 1.1.0 with its null results

- **Question:** port the Aftermarket Report’s Fear & Greed 1.1.0 (NSE-only inputs) onto Data bank inputs, and publish its pre-registered null results next to the gauge.
- **Method:** as specified in the AMR documentation (`FEAR_GREED_RISK_RESULTS.md`), read before porting; the port is checked by reproducing AMR’s published values over their overlap to within rounding.
- **Prediction C8:** the gauge does not predict 1, 3 or 6-month forward returns (as AMR found).

## C9. Do foreign investors chase sector returns?

Written 2026-10-06, before the NSDL sector data was loaded or any of this was computed.

- **Question:** across sectors, do FPIs put money into the sectors that have just done well, and does where they put money say anything about which sectors do well next?
- **Data:** NSDL’s fortnightly sector-wise FPI report (Data bank `nsdl_fpi_fortnightly_sector`, January 2012 on), equity column, rupees. For sector *s* and fortnight *t*, one report gives the opening AUC *A₀*, the fortnight’s net investment *F* and the closing AUC *A₁*; all three come from the same report.
- **Return:** there is no sector index that matches NSDL’s sectors, so the return is the one implied by FPIs’ own holdings (modified Dietz): *R = (A₁ − A₀ − F) / (A₀ + F/2)*. It includes price moves on what FPIs hold, plus noise from corporate actions and companies moving between sectors. **Flow intensity:** *f = F / A₀*.
- **Sectors:** every printed sector except Sovereign, Others and the Grand Total, in a fortnight where its opening equity AUC is at least ₹5,000 crore. NSDL changed its classification on 30 April 2022, so the two eras (to 15 April 2022, and from 30 April 2022) are tested separately and pooled; no fortnight straddles the break.
- **Method:** in each fortnight, the Spearman rank correlation across sectors between *f* at *t* and *R* at *t−1* (chasing), *t* (same fortnight), *t+1* and *t+2* (prediction). The statistic is the mean of these correlations over fortnights. Its t-statistic uses Newey-West errors with 4 lags. A placebo shuffles sector labels within each fortnight 1,000 times.
- **Timing:** NSDL dates FPI trades a day after the exchanges, so fortnight edges are a day out; a fortnight-level test absorbs this. Each report appears a few days after its fortnight ends. A flow at *t* is therefore known only during *t+1*, so *t+2* is the first horizon an outsider could act on, and is reported as such.
- **Prediction C9:** flows correlate positively with the previous fortnight’s sector return (mean correlation above zero, t above 2, pooled). Flows do not predict sector returns at *t+1* or *t+2* (|t| below 2).

## C10. Narrow participation and later returns

Written 2026-10-14, before any of this was computed (owner’s request: the breadth pages said “not tested”).

- **Question:** when the index rises while few stocks take part, have later returns been weaker?
- **Data:** the liquid-universe breadth table (`breadth_stocks_daily.parquet`: share of stocks above their 200-day average, from the first date the universe reaches 200 stocks), the equal-weight minus cap-weight Nifty 500 one-year gap (`breadth_ew_vs_cw.parquet`, from 2006-03), and the Nifty 500 TRI.
- **Signals, at each month-end, from data up to that day:**
  
  - **S1, a narrow rally:** the Nifty 500 TRI is within 5% of its 52-week high **and** the share of stocks above their 200-day average is below its median over all month-ends up to that date (expanding, at least 36 month-ends).
  - **S2, the equal-weight gap:** the Nifty 500 equal-weight minus cap-weight one-year return gap, in points.
- **Forward returns:** the Nifty 500 TRI over the next 3, 6 and 12 months, from the month-end close.
- **Method:**
  
  - **S1:** the mean forward return in flagged months against unflagged months. Significance: 10,000 random circular shifts of the set of flagged months through the sample (keeping their clustering, as in C2 amendment 2); the p-value is the share of shifted sets whose mean forward return is at most the observed one (the claim is “weaker”).
  - **S2:** the slope of the forward 12-month return on the gap, with a Newey-West t-statistic (12 lags) and the range of slopes across the 12 non-overlapping monthly offsets; and median forward returns by tercile of the gap (expanding terciles, known at each date) with the number of independent years in each.
- **Prediction C10:** S1 months have a lower mean 12-month forward return than other months, but the shift p-value is above 0.05; the S2 slope has |t| below 2. I expect narrow markets to look worse afterwards in a few memorable episodes and not reliably.

## C11. Breadth washouts and euphoria, and what followed

Written 2026-10-22, before any forward return was computed for this block. Disclosure: the thresholds below were set after seeing the chart of the share of stocks above the 200-day average itself (`/markets/`), but not any return that followed it.

- **Question:** did a broad washout (very few stocks above their 200-day average) precede better Nifty 500 returns than usual, and did near-universal participation precede worse ones?
- **Data:** `pct_above_200dma` from `breadth_stocks_daily.parquet` (liquid survivorship-free universe; sample from the first date the 60-session median universe reaches 200 stocks, as amended for C2); Nifty 500 TRI.
- **Signals:** a *washout* starts on the first session the reading closes below 15% after a session at or above 15%; a *euphoria* starts on the first session it closes above 85% after one at or below 85%. A new signal of the same kind needs 126 sessions since the last.
- **Timing:** the reading uses closing prices of day *t*, so returns run from the close of *t + 1* (no same-day trading on the signal).
- **Method:** forward 63, 126 and 252-session TRI returns after each signal, against the unconditional mean over the same sample. Significance by 10,000 circular shifts of the whole set of signal dates (as amended for C2), one-sided in the predicted direction. Also, descriptively: the median forward 252-session return by quintile of the daily reading, with the effective number of independent years (sessions ÷ 252) in each quintile.
- **Prediction C11:** fewer than 12 washouts and fewer than 12 euphoria signals. After washouts the mean 252-session return is above the unconditional mean, but p is above 0.05 at 63 sessions. After euphoria, no reliable difference at any horizon (p above 0.05). With this few signals and three horizons on each side, nothing here will be called a finding unless p is below 0.01.

## C12. Do NSE’s indices hold up after launch?

Written 2026-10-22, before any pre- or post-launch return was computed.

- **Question:** NSE publishes years of back-calculated history for most indices, computed before the index existed. Rules-based indices are usually designed with that history in view. Did strategy, theme and sector indices beat the Nifty 500 by less after launch than in their back-calculated years?
- **Data:** daily total-return levels (`index_tri.parquet`), launch and base dates frozen from NSE factsheets (`compute/trend_launch_dates.csv`; the 2026-04-30 report-month artefact and missing dates excluded), Nifty 500 TRI as the benchmark.
- **Sample:** every index in the strategy, thematic and sectoral groups with at least three years of history both before and after its launch date. Broad-market indices meeting the same rule are reported separately as a control.
- **Measure:** annualised excess log return over the Nifty 500 TRI, from the base date to the launch date (*pre*) and from the launch date to the latest date (*post*). Decay = post − pre, per index.
- **Tests:** per group, the median decay, the share of indices with post below pre, and a Wilcoxon signed-rank test of post against pre. Indices overlap heavily in time and holdings, so they are not independent: the p-value is a guide, and the time-series check is a pooled average excess return by calendar month, pre-launch months against post-launch months, with Newey-West errors (12 lags).
- **Amendment, same day, before any result:** scipy is not in the environment, so the Wilcoxon signed-rank test becomes a paired sign-flip permutation test of the mean decay (100,000 draws, two-sided).
- **Prediction C12:** strategy indices beat the Nifty 500 by more than 2% a year on average before launch, and by less than half that after; more than 60% show decay. Sector and theme indices show a smaller decay. Broad indices show none worth naming (median decay within ±1% a year).

*Numbering: these two blocks were written on 2026-10-22 as C10 and C11 in a working copy that did not yet have the C10 above; renumbered C11 and C12 when merged, with no change to their content.*

## C20. Mid caps’ valuation against large caps, and what followed

Written 2026-10-22 by the valuation session, before any forward return below was computed.

- **What I had seen beforehand:** the level of each size index’s P/B against the Nifty 50’s (Midcap 100 at about 1.4 times today, its lowest about 0.5 in early 2009), and that NSE’s P/B has a one-day break on 29 September 2023 (implied book value up 25% for the Nifty 50, 0.5% for the Midcap 100, with prices flat). No forward return of any size index on any valuation signal. General knowledge of the episodes (mid caps’ 2008 and 2018 falls, their 2020–24 run) is common to anyone who follows Indian markets.
- **Question:** when mid caps were dear against the Nifty 50, did they lag it over the next three years?
- **Data:** NSE’s daily P/B for the Nifty 50 and the Nifty Midcap 100 (from July 2005, its launch), on one basis across the 2023 break: book value before 29 September 2023 is scaled by each index’s ratio of implied book value that day to the day before (`pb_adjusted`). Total-return indices from `asset_tri_daily.parquet`. Completed month-ends only.
- **Signal:** log(Midcap 100 P/B ÷ Nifty 50 P/B) at the end of month t−1.
- **Target:** the annualised log return of the Midcap 100 TRI minus the Nifty 50 TRI over the 36 months from the end of month t, in percent a year.
- **Test:** the D1 machinery (`compute/valuation_forecast.py`), unchanged: overlapping OLS slope with Hodrick 1B errors; a bootstrap of 10,000 samples under no predictability that keeps the signal’s persistence and its correlation with returns (Stambaugh bias), one-sided for a negative slope; the bias-corrected slope; out-of-sample R² against the historical-mean forecast using only windows closed at each date (at least 60), with Clark–West; the range of non-overlapping slopes. Seed 20261022.
- **Hypothesis C20 (one test):** passes if the bootstrap p is below 0.05, the bias-corrected slope is negative and the OOS R² is positive.
- **Reported, not tested:** 1- and 5-year horizons; the Nifty Smallcap 100 (from March 2011) and the Nifty Next 50 (from 1999) against the Nifty 50 on the same rules.
- **Prediction C20:** the slope is negative (dearer mid caps, weaker relative returns), but C20 does not pass: about six independent three-year periods are too few to tell a slope from the bias of a persistent signal.

## C21. Sectors cheap or dear against their own past, and the next year

Written 2026-10-22, before any of this was computed.

- **What I had seen beforehand:** today’s P/E and P/B percentiles by index on the valuation page. No sector’s forward return on any valuation signal.
- **Question:** did sectors that were cheap against their own history beat sectors that were dear over the next twelve months?
- **Universe:** ten NSE sector indices with valuation and total-return history from their launch: Bank, IT, FMCG, Pharma, Energy, Metal, Auto, Realty, Infrastructure and PSU Bank. Financial Services is left out because it largely repeats Bank. All data used is from after each index’s launch, so none of it is back-calculated.
- **Signal, at each month-end t:** the percentile of the sector’s month-end P/B (`pb_adjusted`) among its own month-end readings up to t, needing at least 36 months of history. P/B rather than P/E: the 2021 earnings switch cannot be bridged for Auto, Metal or Realty, and P/B carries no such break once the 2023 one is adjusted.
- **Portfolios:** at each month-end with at least six eligible sectors, the cheapest third and the dearest third by count (⌊n/3⌋ sectors each), equal-weighted. The return is each sector’s TRI from the next session’s close (the signal is known only after the month-end close) for twelve months.
- **Statistic:** the mean over start months of the cheap-minus-dear twelve-month return. Newey-West t with 12 lags; a placebo that shuffles the signal across sectors within each month (2,000 shuffles, seed 20261022); p is the share of shuffles with a mean at least as large as observed (one-sided). Independent years are printed beside it.
- **Reported, not tested:** the same spread in months when the cross-sector dispersion of the signal (interquartile range of the percentiles) is above its expanding median, against months when it is below.
- **Hypothesis C21:** passes if the mean spread is positive and the placebo p is below 0.05.
- **Prediction C21:** the mean spread is positive (cheap sectors did a little better), and the placebo p is above 0.05.

## C22. Market value to GDP and the next five years

Written 2026-10-22, before any of this was computed.

- **What I had seen beforehand:** the series’ level (25% of GDP in 1999, about 135% now) and its peaks, as charted on the breadth page. No forward return on it.
- **Question:** has the market value of NSE-listed companies against GDP said anything about the Nifty 50’s real return over the next five years?
- **Data:** `density_mcap_gdp.parquet` (NSE month-end market value against the previous complete financial year’s nominal GDP, latest vintage, from April 1999); the Nifty 50 TRI and CPI exactly as in D1.
- **Signals, at the end of month t−1:** (a) log market value to GDP; (b) its log ratio to its own trailing 120-month average (needs 120 months, so from 2009), which removes the slow rise from more companies listing.
- **Target and test:** the D1 machinery at five years, real, as in C20; one-sided for a negative slope. Seed 20261022.
- **Hypothesis C22 (two tests, a and b; Bonferroni 0.025 each):** each passes if its bootstrap p is below 0.025, the bias-corrected slope is negative and the OOS R² is positive.
- **Prediction C22:** neither passes. Signal (a) trends upward with listings, so its slope will be weak or the wrong sign; signal (b) will have a negative slope on too few independent periods (fewer than three from 2009) to call.

## C30–C32. Flows scaled by market size: persistence, coincidence with falls, and what followed heavy foreign selling

Written 2026-10-22 by the flows-page session, before any of the three was computed. What I had seen beforehand: the C3 results above; the published charts of monthly FPI and DII flows in crore and of their rolling twelve-month sums (so I know FPIs sold heavily in 2024-26 while DIIs bought a record amount); no forward return, autocorrelation or same-month return comparison on scaled flows.

**Common data and scaling (all three blocks):**

- **FPI:** NSDL’s fortnightly equity net investment (primary and secondary market, `fpi_totals.parquet`, INR, Grand Total), summed into calendar months, January 2012 on.
- **DII:** exchange-reported daily DII net purchases (`flows_daily.parquet`), summed into calendar months, April 2007 on.
- **Market size:** total market capitalisation of NSE-listed companies at each quarter-end, implied by PRIME’s quarterly ownership table (`prime_shareholding_history`, all NSE: FII value ÷ FII share by value; it matches the filings-based aggregate within 1% where both exist). Between quarter-ends it is rolled forward with the Nifty 500 TRI from the latest quarter-end. A month’s flow is divided by the market cap at the end of the **previous** month, so the denominator is known before the flow. Unit: basis points of market cap.
- **Returns:** Nifty 500 TRI.

### C30. Do flows persist?

- **Question:** is a month of foreign (or domestic institutional) net selling more likely to be followed by another, and by how much?
- **Rule:** for each flow, the lag-1 autocorrelation of monthly scaled flow; the share of months whose sign matches the previous month’s, against the share expected if signs were independent with the same base rate (p² + (1−p)²).
- **Test:** p-values from 10,000 random permutations of the months (which destroy any ordering); the same statistics on each half of the sample; the length of the current same-sign streak and where it ranks among all completed streaks.
- **Prediction C30:** FPI scaled flows are positively autocorrelated (lag-1 above 0.2, permutation p below 0.05) and their sign persists more than independence implies. DII flows are more persistent than FPI flows, and more so in the second half of the sample (the SIP era), where DIIs were net buyers in most months.

### C31. Did foreign selling coincide with falls more than chance?

- **Question:** in months when FPIs sold, did the market fall more than in other months, and did domestic buying that matched or exceeded the foreign selling go with better months?
- **Rule (same month, descriptive):** the correlation of monthly scaled FPI flow with the same month’s Nifty 500 TRI return; the mean return in FPI net-selling months against net-buying months; among FPI-selling months, the mean return when DII net buying was at least as large as the FPI selling (“absorbed”) against when it was not.
- **Test:** 10,000 random permutations of the flow series against the return series (two-sided for the correlation and the selling/buying gap; for the absorbed/not gap, two-sided as well). Sample: January 2012 on, where both series exist.
- **Prediction C31:** the same-month correlation is positive and clearly significant (above 0.3, p below 0.01), and FPI-selling months average lower returns. Absorption does not mark better months: the absorbed/not gap has p above 0.05, because DIIs buy most heavily into falls (C3), so heavy domestic buying goes with weak months as often as with firm ones.
- **Caveat stated in advance:** same-month association says nothing about which moved which; NSDL counts primary-market buying (IPOs, QIPs) that the exchange DII series does not.

### C32. What followed heavy foreign selling?

- **Question:** after twelve months of unusually heavy FPI selling (scaled by market cap), were the next twelve months’ returns better or worse than usual?
- **Signal:** at each month-end, the trailing twelve-month FPI flow in basis points of market cap (each month scaled as above). A month is **flagged** when that figure is in the bottom fifth of all month-end values up to and including that date (expanding, at least 36 month-ends of history, so the first possible flag is December 2015).
- **Timing:** NSDL’s report for a fortnight appears a few days after it ends, so each forward return starts at the close of the fifth trading session of the following month and runs twelve months (to the close of the fifth session twelve months later).
- **Test:** the mean forward 12-month return in flagged months against unflagged months, with a p-value from 10,000 random circular shifts of the set of flagged months (as C10, keeping their clustering), two-sided; and the slope of the forward return on the trailing scaled flow, with a Newey-West t-statistic (12 lags) and the range of slopes across the 12 non-overlapping monthly offsets. The effective sample is reported in independent years.
- **Prediction C32:** the popular reading is that heavy foreign selling marks a buying opportunity. I expect flagged months to show a higher mean forward return than other months but with a shift p above 0.05, and the slope to have |t| below 2: about ten independent years cannot separate a few memorable rebounds (2016, 2020, 2023) from chance.

## C40 to C43. IPOs: size, insiders and survivorship (the /ipo/ page)

Written 2026-10-22, before any return in these blocks was computed. What I had seen beforehand: everything already on /ipo/ (the H1 to H3 results in `ipo_spec.md`; buy-and-hold returns against Nifty 500, Nifty500 Equal Weight and Smallcap 250 by horizon and by listing year; returns from the issue price grouped by the first day). I had not seen any return grouped by company size or promoter holding, any return against a size-matched portfolio, nor any C4 number. Before writing this I checked only coverage: share counts in the XBRL shareholding data reach back to listing only for 2025 and later issues; promoter holdings near the listing date exist for listings from 2021; AMFI’s half-yearly average market-capitalisation lists start with July–December 2017.

Common rules for C40 to C42:

- **Sample:** NSE mainboard listings from `ipo_listings.parquet` (the `ipo_spec.md` universe: fresh ISIN, BSE-first listings excluded, delisted stocks kept).
- **Size without look-ahead.** AMFI’s list for a half-year gives each company’s average market capitalisation over that half-year. Dividing it by the stock’s mean NSE close over the same sessions gives its share count, which does not depend on how the price moved. Market value on any later date is that share count times the close (adjusted prices on both sides, so splits and bonuses cancel). A list counts as public 15 days after its period ends.
  
  - **An IPO’s size at listing** = shares implied by the first AMFI list whose period contains the listing date, times the listing-day close. Only the share count comes from the later list; no later price enters.
  - **Validation before use:** for 2025–26 listings with an XBRL share count filed within 45 days of listing, the implied count must be within 10% of the filed count for at least 80% of issues. If not, the method is amended (logged) before any return is computed.
- **Returns:** price returns from the adjusted panel, horizons in NSE sessions as in `ipo_spec.md` (250 = 1 year, 750 = 3 years). A stock that stops trading exits at its last close.

### C40. IPOs against a size-matched portfolio

- **Question:** once an IPO is compared with seasoned companies of the same size, does it still trail, beat, or match them?
- **Peers:** for each IPO, the 10 companies nearest in log market value on the last session before listing, from the latest AMFI list public by then. Peers must have traded on NSE for at least 36 months before the listing (seasoned), trade on the session before it, and not be the IPO. Peer market value = its implied share count (from that list) times its adjusted close on the session before the listing.
- **Sample:** listings from 2018-01-16 (the first AMFI list, for July–December 2017, is public on 2018-01-15) whose own size can be measured.
- **Primary test:** a calendar-time version of H1. Each month, the equal-weighted return of every sample IPO in months 1 to 36 after its listing month, minus the equal-weighted return of the same IPOs’ peer baskets over the same month (each basket an equal-weighted average of its peers’ returns that month). Months need at least 10 IPOs. Statistic: the mean monthly difference with a Newey-West t (3 lags).
- **Secondary (descriptive):** per issue, the 1-year and 3-year buy-and-hold abnormal return against the peer basket (equal-weighted buy-and-hold from the listing-day close to the same end date): n, mean, median, share beating. The H2 Spearman (listing gain against the 1-year BHAR) repeated with size-matched BHARs.
- **Prediction C40:** no difference: the mean monthly gap is within ±0.5 points and |t| < 2. The median 3-year BHAR against size-matched peers is negative and under half of IPOs beat their peers, as against Nifty 500.

### C41. Big and small IPOs

- **Question:** do small IPOs do worse afterwards than large ones, even against companies of their own size?
- **Sample:** the C40 sample with a complete 1-year horizon.
- **Test:** Spearman correlation between log size at listing and the 1-year BHAR against the size-matched peers; t by the normal approximation, and a 95% interval from 2,000 bootstrap resamples of listing months (issues in the same month resampled together).
- **Descriptive:** by size tercile (cut on the sample): n, median size, median listing gain, median 1-year return from the issue price, median and mean 1-year BHAR against peers and against Nifty 500, and the share beating peers.
- **Prediction C41:** a positive correlation (small IPOs worse), with |t| < 2.

### C42. The promoter’s stake after listing

- **Question:** do IPOs where the promoter keeps less of the company (often private-equity exits or professionally run firms) do worse afterwards?
- **Data:** promoter and promoter-group holding from the first shareholding filing (`nse_shareholding_summary`) dated within 45 days before to 120 days after the listing date. Listings from 2021-01-01 (earlier listings mostly lack a filing near listing).
- **Test:** Spearman correlation between the promoter stake and the 1-year BHAR against Nifty 500; t by the normal approximation and the listing-month bootstrap interval as in C41.
- **Descriptive:** by stake group (no promoter, under 50%, 50 to 70%, 70% and above): n, median listing gain, median 1-year return from the issue price, median 1-year BHAR against Nifty 500 and, where measured, against size-matched peers.
- **Prediction C42:** a positive correlation (low promoter stake, worse later), with |t| < 2.

C40, C41 and C42 are three primary tests: a p-value must be under 0.0167 (Bonferroni) to count.

### C43. Survivorship: how much do failures move the results?

- **Question:** the IPO results exit a stock that stops trading at its last close. If such stocks are instead treated as total losses, do the conclusions change?
- **Method:** list every mainboard listing whose last NSE trade is more than 30 days before the panel’s last session, with its last trade date and return to it. Re-run H1 (Nifty 500 intercept) and the 3-year mean and median BHAR against Nifty 500 with each such stock given a −100% return in the month after its last trade (and at its exit in buy-and-hold terms). Also report the issues excluded from the universe because their symbol never traded within 15 sessions, with reasons.
- **Prediction C43:** the H1 alpha moves by less than 1 point a year and the 3-year mean BHAR by less than 2 points, because very few mainboard IPOs since 2016 stopped trading.

### C4, data and timing (amendment written 2026-10-22, before C4 was computed)

- **IPO money:** mainboard plus SME IPO amounts (SEBI, `ipo_fundraising_monthly.parquet`), summed over the trailing 12 months, from 2013-03 (12 months of data from 2012-04).
- **Market capitalisation:** NSE’s month-end market value of listed companies (`density_mcap_gdp.parquet`, `mcap_cr`). The signal is the trailing 12-month IPO money divided by the month-end market value.
- **Timing:** SEBI’s monthly figures appear about a month later, so a signal for month *m* is used from the end of month *m*+1; forward 12 and 36-month Nifty 500 TRI returns run from that month-end.
- **Statistics:** as the common rules: slope, R², Newey-West t with *h* lags, the range of slopes across the *h* non-overlapping offsets, the number of independent periods, and median forward returns by expanding tercile of the signal.

## C50–C53. Mutual funds: persistence, new launches and return-chasing (funds pages)

Written 2026-10-22 for `/funds/` and `/funds/passive/`, before any forward return below was computed. What I had seen: the published charts of NFO collections, category assets and flows (levels, never set against later returns), and India’s market path in broad strokes (the 2020 crash, the 2021 rally, the 2024 peak and the correction after it). The descriptive scorecard of active funds against their benchmarks (share beating over trailing windows) is not a forward test and is not pre-registered; it shares C50’s data and rules.

Common data and rules for C50–C51:

- **Funds:** every open-ended active equity scheme in AMFI’s NAV history (all eras, closed and merged schemes kept) with exactly one direct and one regular growth plan (the pairing in `mf_realised_cost.plan_pairs`, with dividend-yield funds kept: the word “dividend” in their name is not an option). Direct plans are the headline (the fund’s own cost only; index funds are compared on the same footing); regular plans are reported beside them. Index funds, ETFs, FoFs, arbitrage and hybrids are out.
- **Category and benchmark:** the scheme’s last AMFI category (AMFI’s NAV file prints the current label, not the label at each date; disclosed). Tier-1 total-return benchmarks: large cap → Nifty 100 TRI; mid cap → Nifty Midcap 150 TRI; small cap → Nifty Smallcap 250 TRI; large and mid cap → Nifty LargeMidcap 250 TRI; multi cap → Nifty 500 Multicap 50:25:25 TRI (back-calculated before December 2020, used here only as a hurdle); flexi cap, ELSS, focused, value, contra and dividend yield → Nifty 500 TRI; sectoral and thematic → Nifty 500 TRI (the broad market the investor could have held instead; their own sector indices are not used). Schemes whose last label is a pre-2018 one (“Growth”) have no category and are counted only in the all-funds totals, against the Nifty 500 TRI.
- **Returns:** month-end NAV to month-end NAV, annualised; the benchmark over the same dates.
- **Closed and merged funds:** a fund alive at a window’s start that stops before its end counts as **not beating** (the SPIVA convention); as a check, the same fund judged on its return to its last NAV against the benchmark over the same span.

### C50. Do last period’s winners keep winning?

- **Question:** among active equity funds in the same category, do those with the best three years go on to beat their benchmark, or stay in the top quarter, over the next three years?
- **Formation:** each December from 2015 to 2022 (eight cohorts, evaluation ending December 2018 to December 2025): funds alive with three full years of direct-plan NAVs; excess return = fund CAGR minus benchmark CAGR over the three years; quartiles within category (categories with fewer than 8 funds at formation are dropped from the quartile test).
- **Evaluation:** the next three years’ excess return; a fund that dies counts in the bottom quartile and as not beating.
- **Tests:** (1) the share of formation top-quartile funds that are top-quartile again, against 25% under no persistence; (2) the share of top-quartile and of bottom-quartile funds that beat their benchmark next; (3) the Spearman rank correlation within category between formation and evaluation excess returns, averaged over categories in a cohort. Headline: the three non-overlapping cohorts (December 2015, 2018, 2021), significance by 10,000 within-category shuffles of the evaluation ranks; all eight cohorts as the overlapping robustness check.
- **Prediction C50:** persistence is weak. Fewer than 35% of top-quartile funds are top-quartile again; the mean rank correlation is between −0.1 and +0.2 and the shuffle p-value on the non-overlapping cohorts is above 0.05.

### C51. What followed new active equity funds?

- **Question:** how did new active equity funds do over their first one, two and three years, against simply holding the Nifty 500, and were sectoral and thematic launches worse?
- **Launches:** direct-plan growth codes of active equity schemes (as above, including sectoral and thematic) whose first NAV is between 1 January 2019 and 30 September 2025 and whose scheme had no earlier plan code (excludes conversions and renamings). Launch date = first NAV date (allotment).
- **Rule:** fund return from the first NAV to the NAV 12, 24 and 36 months later (the nearest earlier NAV), against the Nifty 500 TRI over the same dates; a fund that closes earlier is judged on its return to its last NAV against the TRI over the same span, and flagged. Horizons are computed only where the full window has passed.
- **Test:** the median excess return and the share beating the TRI, for sectoral and thematic and for diversified launches; 95% intervals by a bootstrap that resamples launch quarters (launches cluster in time and share the market’s path).
- **Prediction C51:** over two years, the median new active equity fund lags the Nifty 500 TRI and fewer than half beat it; sectoral and thematic launches lag by more than diversified ones; the interval for diversified launches includes zero.

### C52. Do equity NFOs come at market highs, and what followed?

- **Question:** is money raised by new equity funds highest after the market has risen, and has a hot NFO market been followed by weaker returns?
- **Data:** equity NFO collections per month from AMFI’s monthly report (`density_funds_nfo_monthly`, April 2018 on), active equity only; equity fund AUM at the previous month-end from the same report; Nifty 500 TRI month-ends.
- **Signal:** NFO intensity = the trailing three months’ equity NFO money ÷ equity fund AUM three months earlier. The report for month *t* is out by about the 10th of *t+1*, so the signal is first usable at the end of *t+1*.
- **Tests:** (1) descriptive: the correlation of intensity with the Nifty 500’s trailing 12-month return to *t*; (2) forward: the slope of the Nifty 500 TRI return from the end of *t+1* over the next 12 months on intensity, Newey-West t with 12 lags, the range across the 12 non-overlapping offsets, and the mean forward return in the top third of intensity months (expanding terciles, at least 24 months) against the rest. The 2022 SEBI pause months (no launches allowed) are dropped, not counted as zero.
- **Prediction C52:** intensity rises with the trailing return (correlation above +0.3); the forward slope is negative but |t| < 2, with about seven independent years.

### C53. Do equity fund investors chase category returns?

- **Question:** across equity fund categories, does money flow into the categories that have just done well, and does it say anything about which do well next?
- **Data:** AMFI monthly report, May 2019 on: net flow *F* and month-end AUM *A* for large cap, mid cap, small cap, large and mid cap, flexi and multi cap (one category, as in the overview chart), ELSS, focused, value and contra, and dividend yield. Sectoral and thematic is left out: its contents change with the themes in fashion.
- **Return:** each category’s own realised return implied by its AUM, *R = (A₁ − A₀ − F) / (A₀ + F/2)*, as in C9. **Flow intensity:** *f = F / A₀*.
- **Method:** each month, the Spearman correlation across categories between *f* at *t* and the category’s return over the previous 12 months (*t−12* to *t−1*), and between *f* at *t* and its return over the 12 months from *t+2* (the report for *t* is out during *t+1*). The statistic is the mean correlation over months, Newey-West t with 12 lags; a placebo shuffles category labels within each month 1,000 times.
- **Prediction C53:** flows correlate positively with the past 12 months’ return (mean correlation above zero, t above 2). They do not predict the next 12 months (|t| below 2).

## C60–C63. The economy’s cross-dimension measures: do they track or lead growth? (economy portal front door)

Written 2026-10-24 by the econ-models session, before any of these was computed. Code will be `pipeline/tipsheet/compute/economy_validation.py`; measures as defined in `docs/methods/economy_models.md`.

- **What I had seen beforehand:** the economy breadth series (share of indicators improving, monthly from 2008) and the scorecard’s latest readings, including real GVA and IIP growth as levels. I had not set breadth, the composites or the financial conditions index against any later or concurrent growth figure. The financial conditions index and the composites had not been built. General knowledge of India’s cycle (the 2008-09 slowdown, the 2019 slowdown, the 2020 lockdown, the 2022-23 tightening) is common to anyone who follows the economy.
- **Common rules:** every observation whose signal window or outcome window touches April 2020 to March 2022 is dropped (the lockdowns and their base effects would decide any slope on their own). GVA is real GVA growth on a year earlier, bases linked as on `/macro/`; the old and new bases are uncorrelated where they overlap, so results are also reported for the old base alone (to FY25). Newey-West standard errors with h lags for overlapping windows; non-overlapping subsamples reported beside them. Seed 20261024 wherever resampling is used.

### C60. Does a tightening of financial conditions precede slower industrial growth?

- **Signal:** the change in the India FCI (variant A, equal-weight groups, positive = tighter) over the three months to the end of month t, built only from inputs published by then.
- **Target:** the change in IIP growth (3-month average of growth on a year earlier) from month t to month t+6, the IIP for t+6 being known about six weeks later.
- **Regression:** target on signal, with the IIP growth change over the three months to t as a control. Benchmark: the same regression with the change in RBI’s month-end 10-year less 1-year yield spread in place of the FCI.
- **Pass (one test):** the FCI slope is negative with Newey-West t below −2 at h = 6, and its R² is above the benchmark’s.
- **Reported, not tested:** h = 3 and 12; the mean IIP change after months when the FCI tightened by more than one standard deviation over three months, against all other months.
- **Prediction C60:** the slope is negative but C60 does not pass. IIP growth is noisy, the FCI’s full set of inputs starts only in the 2010s, and there are perhaps eight independent six-month windows a decade.

### C61. Does a tightening of financial conditions precede slower bank credit growth?

- **Signal:** as in C60. **Target:** the change in scheduled commercial banks’ credit growth (merger-adjusted, the month’s last reporting date) from t to t+6. Control: the credit growth change over the three months to t. Bank credit growth is itself an FCI input in no variant (credit growth is held out of the index for this test; if the index as published includes it, C61 uses the index rebuilt without it).
- **Pass:** slope negative, Newey-West t below −2 at h = 6.
- **Prediction C61:** passes more easily than C60. Lending responds to rates with a lag and is smoother than IIP. A coin flip; I expect a negative slope with t near −2.

### C62. Does the activity composite track GVA better than any one indicator?

- **Data:** the monthly activity composite (reference-month version, z against 36 previous readings), averaged over each quarter’s three months; each activity indicator’s own z, averaged the same way; real GVA growth for the quarter.
- **Test:** the correlation of the quarter-average composite with GVA growth, from FY14 (the first year enough themes exist) to the latest quarter, against the best single indicator’s correlation over the same quarters (indicators with at least 80% of those quarters). A block bootstrap (blocks of four quarters, 10,000 draws) gives the interval of the difference.
- **Pass:** the composite’s correlation is above the best single indicator’s, and the bootstrap interval of the difference lies above zero.
- **Prediction C62:** the composite correlates positively (above 0.4). It does not beat the best single indicator by a margin the bootstrap can confirm: IIP or core output is likely to match it.

### C63. Is breadth coincident with growth, or does it lead?

- **Data:** economy breadth (share improving, point-in-time) at each quarter’s last month-end; the change in real GVA growth from the quarter before (acceleration).
- **Test:** correlations at lag 0 (breadth at the end of quarter q against q’s acceleration) and lag 1 (against q+1’s), with block-bootstrap 95% intervals (blocks of four quarters). Breadth at the end of q is known before q’s GDP release, so lag 0 is a nowcast, not a coincidence in the publication sense.
- **Prediction C63:** coincident. The lag-0 correlation is positive with an interval above zero; the lag-1 correlation is smaller and its interval includes zero. If so, the page says breadth tells you about the quarter just ended before GDP does, not about the next one.

## C80–C82. Money: how fast do repo moves reach loans and deposits? (`/economy/money/transmission/`)

Written 2026-10-24 by the econ-money session, before any pass-through below was computed. Code will be `pipeline/tipsheet/compute/economy_money.py`. What I had seen beforehand: the last three rows of `rbi_transmission_monthly` (rate on new loans 8.52% in July and 8.61% in August 2026, on new term deposits 5.85% and 5.67%, repo 5.25%); the full list of repo moves since 2010 (`rbi_rate_moves`), including the 25 bp hike of October 2026; and RBI’s published view (Monetary Policy Reports, 2020-2025) that linking new floating-rate retail and small-business loans to an external benchmark from 1 October 2019 sped up transmission. I had not computed any change in a lending or deposit rate over any cycle.

- **Data:** RBI’s monthly weighted average lending rate on fresh rupee loans (WALR fresh, from late 2014) and on outstanding loans (from 2012), weighted average domestic term deposit rate on fresh deposits (from December 2020) and on outstanding deposits (from 2013), the median one-year MCLR (from 2016), all as in `rbi_transmission_monthly`; repo moves from `rbi_rate_moves`.
- **Cycles:** a maximal run of repo moves of one sign. First-move month *m1*, last-move month *mL*, the cycle’s repo change *R* (bp, signed). Baseline month *b* = the month before *m1*. The window ends *w* = the month before the next cycle’s first move, or the latest data. Nothing after *w* is used, so one cycle never borrows the next one’s moves.
- **Measures, for each rate X with a value at *b*:** pass-through at *h* months after the last move = (X(*mL*+*h*) − X(*b*)) / *R*, for *h* = 0, 3, 6, 12, blank when *mL*+*h* is after *w* or X is missing; and months to half, the first month *t* ≥ *m1* with (X(*t*) − X(*b*)) / *R* ≥ 0.5, counted in months from *m1* (0 = the month of the first move), reported as “not within the cycle” when it is not reached by *w*.
- **C80, external benchmarks:** cycles whose first move came after 1 October 2019 (the 2022-23 tightening, the 2025 easing) against earlier cycles with WALR fresh at baseline (the 2015-17 easing, the 2018 tightening, the 2019-20 easing). **Prediction:** WALR-fresh pass-through six months after the last move is higher in each of the later cycles than in each earlier cycle that has a value. With two cycles against at most three, this is a description, not a test with a p-value; it fails if any earlier cycle beats any later one.
- **C81, margins:** fresh deposits against fresh loans, in the two cycles where both exist. **Prediction:** in the 2022-23 tightening, fresh deposit rates had passed through more than fresh loan rates six months after the last move (banks paid up for deposits); in the 2025 easing, fresh deposit rates had passed through less than fresh loan rates at the latest month in the window (margins squeezed). Each half graded on its own.
- **C82, the October 2026 hike (a forecast, graded when data arrive):** baseline September 2026. **Prediction:** by the April 2027 figure (six months after the move), WALR fresh is at least 12.5 bp (half the hike) above its September 2026 level, and WADTDR fresh is up by less than WALR fresh. If the RBI moves again before then, the comparison is against the cumulative repo change and the cycle rules above.
- **Reported, not graded:** outstanding-loan and deposit pass-through, and MCLR, for every cycle; how the average lending rate on new loans moves with loan mix as well as pricing (a caveat on the page).

## C70–C71. Prices: does underlying inflation say where headline is going, and do households know? (`/economy/prices/`)

Written 2026-10-24 by the econ-prices session, before any measure below was computed. What I had seen: the latest headline (5.54%, September 2026) and rebuilt core ex food and fuel (4.60%) in `macro_cpi_monthly`; the latest household survey reading of current inflation (8.37%, September 2026); and Indian inflation’s broad path (high in 2013, low in 2017 and 2019, high again in 2020 and 2022, very low in 2025). I had not computed any trimmed mean, median, seasonal factor, forecast error or survey error.

### C70. Do underlying measures predict headline a year ahead better than core ex food and fuel?

- **Target:** headline CPI (combined) year-on-year at *t*+12, as published (base 2012 to December 2025, base 2024 from January 2026; the convention in `docs/methods/macro.md`).
- **Forecasts made at *t*, each a measure’s own value used as the forecast of headline a year later** (the “attractor” use RBI’s core studies test):
  
  1. headline at *t* (random walk, the floor);
  2. **core ex food and fuel** at *t* (the benchmark named in the question);
  3. weighted median of the 23 base-2012 components’ year-on-year rates (12 food subgroups, pan and tobacco, clothing, footwear, housing, fuel and light, and the six miscellaneous subgroups; MoSPI’s published combined weights; housing filled from urban where the combined is blank, as in `macro.md`);
  4. weighted trimmed means of the same distribution, 10% and 25% of weight cut from each tail;
  5. the long-history **ecumenical median**: the median of headline and core at 3-month and 6-month seasonally adjusted annualised rates and 12-month rates, and measures 3–4 at 12 months, each put on headline’s footing by subtracting its mean gap to headline over the months before *t* (expanding; entered unadjusted until 36 months of gap exist).
- **Seasonal adjustment** for the 3- and 6-month rates: for each calendar month, the mean of that month’s log change minus the mean log change of its year, over the eight calendar years before *t*’s year (at least five), April 2020 to March 2021 excluded. Concurrent: no factor uses data after the year before *t*.
- **Origins:** January 2017 to September 2025 (105 months; targets to September 2026), about nine independent years. Current-vintage data, labelled pseudo-real-time (MoSPI revises CPI only provisional to final, by hundredths).
- **Test:** RMSE and MAE of each; Diebold-Mariano on squared errors against core, Newey-West with 12 lags and the Harvey-Leybourne-Newbold correction, two-sided. Sensitivity: origins in April 2019 to March 2021 dropped (targets or inputs in the Covid imputation months).
- **Prediction C70:** core beats the headline random walk (lower RMSE). Following George et al. (RBI Bulletin, February 2024), no underlying measure beats core by a margin the test can tell from chance: every |DM t| is below 2, and the ecumenical median’s RMSE is within 10% of core’s. The trimmed means do no better than core.
- **The 358-item measures on base 2024** (from January 2026) cannot be tested before targets exist from January 2027; they are reported, not graded, until three years of targets exist.

### C71. Do households’ inflation expectations predict inflation?

- **Data:** RBI’s Inflation Expectations Survey of Households, median expected inflation one year ahead, every round from the survey month, via India Data Hub (`HHIEINFMED12T`); realised headline CPI year-on-year twelve months after the survey month (as in C70), rounds whose target month is January 2013 or later and has been published.
- **Comparison:** the latest headline print known at the survey month (the previous month’s CPI, published around the 12th).
- **Tests:** (1) bias: mean of expectation minus realised, Newey-West t with lags equal to the rounds in a year; (2) accuracy: MAE and RMSE of the expectation against those of the latest print, Diebold-Mariano as in C70.
- **Prediction C71:** households overpredict by more than 2 percentage points on average (t above 2), and their expectation is a worse forecast than the latest print (higher RMSE), with the DM test rejecting equal accuracy at 5%.

## C64–C65. The GDP/GVA nowcast (economy portal, Growth): the live score and the backtest

Written 2026-10-24 by the econ-models session, before any bridge equation was estimated, any backtest error computed or any nowcast made. Code will be `pipeline/tipsheet/compute/economy_nowcast.py`.

- **What I had seen beforehand:** published real GVA and GDP growth by sector to Q1 FY27 on both bases (as levels and growth rates, through the macro tables and the round-1 scorecard); the three new-base GDP press notes (27 Feb, 5 Jun and 31 Aug 2026), including that Q1 FY26 manufacturing growth was revised from 10.4% to 8.3% between June and August; the RBI’s October 2026 projection of 7.2% for Q2 FY27 GDP growth. The monthly indicators’ latest readings (round-1 scorecard). Not seen: any regression of sector growth on indicators, any backtest error, any Q2 FY27 estimate.

**The model (M1, bottom-up bridge).**

- **Targets:** real GVA growth on a year earlier for the eight NSO sectors; headline GVA growth as their weighted sum, each sector weighted by its share of real GVA in the same quarter a year earlier (the identity behind MoSPI’s contributions).
- **Indicators**, each as the quarter’s average of monthly growth on a year earlier over the months available: mining = IIP mining; manufacturing = IIP manufacturing; utilities = IIP electricity; construction = mean of cement and steel output growth (core industries); trade, hotels and transport = mean of whichever of domestic air passengers, goods-vehicle registrations, e-way bills, major-port cargo and diesel consumption are available; finance and real estate = mean of real bank credit and real bank deposit growth (nominal growth less CPI inflation). Agriculture and public administration have no monthly indicator held: each is forecast as the mean of its own last four published quarters.
- **Bridge, per indicator sector:** growth = a + c·(new base) + b·indicator, OLS. Training sample: the old-base (2011-12) series to Q4 FY23 and the new-base (2022-23) series from Q1 FY24, latest vintage; quarters Q1 FY21 to Q1 FY23 are dropped (lockdowns and their base effects).
- **GDP:** the GVA nowcast plus the mean gap between GDP and GVA growth over the last four published quarters.
- **Intervals:** σ = the larger of the old-base and new-base backtest RMSEs (C65) for the information set in use; 70% interval ±1.04σ, 90% ±1.645σ (normal quantiles), for GVA and GDP separately.
- **Information sets:** “two months” (months 1–2 of the quarter’s indicators) and “three months” (all three). The live nowcast uses whatever has been released on the day.

### C65. Does the bridge beat naive benchmarks in a pseudo-real-time backtest?

- **Label:** pseudo-real-time, latest vintage (no first prints exist on either base before 2026, except the three reconstructed new-base quarters).
- **Design:** expanding window. For each evaluation quarter q, the bridges and benchmarks are estimated on target quarters up to q−1 (published before q’s nowcast would be made). (a) Old base only (no dummy): evaluation Q1 FY17–Q4 FY20 and Q2 FY23–Q2 FY26. (b) New base (spliced training with the dummy): evaluation Q1 FY25–Q1 FY27.
- **Benchmarks:** random walk (last quarter’s GVA growth); AR(1) on headline GVA growth over the same training quarters; the mean of the last four quarters.
- **Metrics:** RMSE and MAE of headline GVA growth, both information sets; Diebold–Mariano on squared errors (h = 1) against each benchmark.
- **Pass (one test):** with three months of data, the bridge’s RMSE is below the random walk’s and the AR(1)‘s in both backtests, and the Diebold–Mariano p against the random walk is below 0.10 in the old-base backtest.
- **Prediction C65:** the bridge beats the random walk in the old-base backtest by a modest margin, but the Diebold–Mariano p is above 0.10, so C65 does not pass. In the new-base backtest (9 quarters) nothing can be told apart.

### C64. The live score: Q2 FY27 and every quarter after

- **The scored nowcast** is the last one issued before MoSPI’s release (30 November 2026, 16:00 IST), as logged with its issue time in the bundle’s history. Earlier estimates are kept and shown but not scored.
- **Targets:** the first print of real GVA growth and real GDP growth for Q2 FY27 in MoSPI’s press note.
- **Benchmarks:** random walk (the Q1 FY27 first print: GVA 8.2%, GDP 7.8%); AR(1) on headline growth through Q1 FY27; for GDP, the RBI’s projection in the latest policy statement before the release (7 October 2026: 7.2%); the RBI SPF median for the quarter if econ-data has ingested it before the release.
- **Metric:** absolute error, ours against each benchmark; whether the first print falls in our 70% and 90% intervals.
- **Rule:** one quarter decides nothing and is reported as a single observation, never as a pass or fail. After eight scored quarters: RMSE against each benchmark, Diebold–Mariano, and interval coverage against the stated 70% and 90%.
- **Prediction C64 (Q2 FY27):** the GVA first print falls inside our 70% interval, and our absolute error on GDP is smaller than the random walk’s.

### C66. A challenger to the bridge: indicator growth plus its recent gap

Written 2026-10-24 by the econ-models session, before the 28 October IIP release and before the challenger was computed in any form. At the lead’s request: the C64 model’s sector slopes (manufacturing 1.44 times IIP growth, pooled across both bases, with a new-base intercept resting on 13 quarters) drive most of its distance from the RBI’s projection.

- **What I had seen beforehand:** everything listed for C64–C65, plus the C65 backtest results (the bridge did not beat the random walk or the AR(1) on the old base; it did on the new base over 9 quarters, p 0.21) and the first live nowcast (GVA 9.1%, GDP 9.0%; manufacturing contributing 2.3 pp at a 14.2% sector growth). I have not computed any version of the challenger.
- **The challenger:** for each sector with an indicator, sector growth = the quarter’s indicator growth + the mean of (sector growth − indicator growth) over the last four published quarters (same base; quarters Q1 FY21 to Q1 FY23 skipped). The slope is fixed at 1 and nothing is fitted. Agriculture, public administration, weights, the GDP step, the information sets and the interval rule are as in C64, with the challenger’s own backtest errors.
- **Scoring:** the challenger is published beside the C64 model from its first issue and scored by the C64 rules on the same quarters; its backtest by the C65 rules (reported, not a second test). The C64 model is not changed before the 30 November print.
- **Prediction C66:** in the old-base backtest, the challenger’s RMSE is below the bridge’s (a four-quarter gap follows the old base’s drift away from its indicators). For Q2 FY27, the challenger’s GVA nowcast is below the bridge’s.

## C85–C87. Households and jobs: do surveys and payment flows track what they are read as? (`/economy/households/`, `/economy/jobs/`)

Written 2026-10-24 by the econ-households session, before any correlation, regression or forecast error below was computed. Code will be `pipeline/tipsheet/compute/economy_households.py` and `economy_jobs.py`.

- **What I had seen beforehand:** the latest readings only: RBI’s urban consumer confidence current situation index (87.9, September 2026) and rural (90.6); the urban survey’s net response on employment now (−21.6); PLFS monthly unemployment (5.0%, August 2026) and quarterly urban unemployment from 2018 (9.7% in Apr–Jun 2018, 6.7% in Apr–Jun 2026); MGNREGA’s persondays by financial year (363 crore in 2021-22, 228 crore in 2025-26); the round-3 Growth pages (real household consumption growth by quarter). I had not computed any relationship between these series, nor looked at EPFO, ESIC or NACH growth rates.

### C85. Does urban consumer confidence say anything about household consumption that last quarter’s figure does not?

- **Data:** the current situation index (CSI) of RBI’s urban Consumer Confidence Survey, each round dated by its survey month: RBI’s survey workbook (`rbi_survey_tables`, survey `uccs`, Table 9) where it is held, India Data Hub’s copy of the same RBI figure (`HHCCCURIDX11T`) for earlier rounds (the two must agree within 0.5 in their overlap, else the test stops). Real private final consumption expenditure (PFCE) growth on a year earlier, quarterly, bases linked as on the Growth pages (`economy_growth_quarterly`, spliced: 2011-12 base to Q4 FY23, 2022-23 base after).
- **Timing:** a quarter’s CSI is the mean of the rounds fielded in that quarter (rounds are fielded in the first fortnight of the survey month). Target: PFCE growth in the same quarter, first published about two months after it ends, so the survey is known well before the target. The benchmark’s input, last quarter’s PFCE growth, is published by the end of the quarter’s second month: the benchmark is given information a little later than the survey’s, which favours the benchmark.
- **Predictor:** the change in the quarter’s CSI over four quarters (sentiment against a year ago, matching growth against a year ago).
- **Sample:** target quarters with a CSI and a PFCE figure from Q3 FY13 on, leaving out Q1 FY21 to Q1 FY23 (lockdowns and their base effects, as on the Growth pages) and any quarter whose four-quarter CSI change reaches into those quarters.
- **Test:** expanding-window out-of-sample forecasts from the 17th usable quarter on. Benchmark (a): random walk, PFCE(q) = PFCE(q−1). Benchmark (b): AR(1), PFCE(q) = a + b·PFCE(q−1), OLS. Model: AR(1) plus c·ΔCSI(q). RMSE of each; Diebold–Mariano on squared errors, model against (b), h = 1, two-sided. Reported alongside: in-sample Spearman correlation of ΔCSI with PFCE growth.
- **Pass:** the model’s RMSE is below both benchmarks’ and the DM p against the AR(1) is below 0.10.
- **Prediction C85:** fails. ΔCSI correlates positively with PFCE growth in sample (Spearman between 0.2 and 0.5), but adds nothing out of sample that the last quarter’s figure does not already carry (RMSE within 5% of the AR(1)‘s, DM p above 0.10).

### C86. Do households’ feelings about jobs move with the measured unemployment rate?

- **Data:** the urban Consumer Confidence Survey’s net response on employment, current perception against a year ago (Table 2, “Current Perception > Net Response”: % saying improved less % saying worsened), each round by survey month. PLFS quarterly urban unemployment rate, current weekly status, persons 15+ (`macro_jobs_quarterly.ur_urban`, the identity-checked series on /macro/), from Apr–Jun 2018.
- **Timing:** a quarter’s survey reading is the mean of the rounds fielded in it; matched to the PLFS quarter of the same months. Both are contemporaneous: this is a tracking test, not a forecast.
- **Measure:** four-quarter changes of each (net response, percentage points; unemployment rate, pp).
- **Sample:** quarters where both changes exist, leaving out any change whose start or end quarter falls in Apr 2020 to Jun 2022. PLFS redesigned its sample from January 2025: the main result uses every remaining quarter; a sensitivity drops changes that span the redesign (end quarters Jan–Mar 2025 to Oct–Dec 2025).
- **Test:** Spearman correlation of the two changes; p from 10,000 random permutations of one series against the other (two-sided). Reported alongside: the Spearman correlation of the levels (no p: both trend).
- **Pass:** Spearman of the changes at or below −0.4 (better feeling, lower unemployment) with permutation p below 0.05.
- **Prediction C86:** fails. The levels correlate negatively (perceptions improved as unemployment fell), but four-quarter changes correlate weakly (|ρ| below 0.4), so households’ sense of the job market does not track the survey’s unemployment rate from quarter to quarter.

### C87. Are NACH bulk credits a usable proxy for formal hiring?

- **Data:** NPCI’s NACH credit statistics (`npci_product_statistics`, product `nach`, tab `nach-credit`): successful transaction value of “other credits” (rows `credit_306__other_credits` plus `credit_156__other_credits` where present), the bulk credits companies use for salaries, pensions, dividends and interest, excluding the government’s direct benefit transfers (the `apbs` and `dbt` rows). Formal hiring: EPFO net new subscribers (India Data Hub `LAEMNETNSB11M`, to July 2025), and as a second measure ESIC new subscribers (`LAEMNESIAL11M`).
- **Measure:** each series’ three-month sum, growth on the same three months a year earlier, monthly.
- **Sample:** months from April 2022 (after the lockdown months and their base effects leave the comparison) to the last month both series have: July 2025 for EPFO, the latest common month for ESIC.
- **Test:** Pearson correlation of the two growth rates. Significance by circular shift: one series is shifted against the other by every offset of at least six months (wrapping round), and p is the share of shifts whose |r| is at least the observed |r|.
- **Pass (EPFO the main test, ESIC reported):** r at or above 0.5 with circular-shift p below 0.05.
- **Prediction C87:** fails for both. NACH other credits grow with payrolls and investment income together; their correlation with net new EPF subscribers is below 0.3, so the page labels them a salary-and-income flow, never a jobs count.

C88 and C89 are held in reserve.

## Order of work

C1 first (data ready, small), then C2 and C3 (data ready), C4, C5 (after lane 1’s leverage tables settle), C8 (needs the AMR spec), and C7 (largest; needs a data request for shares outstanding and fundamentals).

## Amendments

**2026-10-02, before C2 and C3 were computed (data constraints found while wiring them):**

1. **C2 sample:** the bhavcopy breadth universe holds fewer than 200 stocks a day before 2004 (9 in 1995, about 100 in 2000), so an advance/decline ratio there is not a market breadth measure. The C2 sample starts on the first day the 60-session median universe reaches 200 stocks. The prediction’s “since 1995” becomes “since that date”.
2. **C2 significance:** the block bootstrap of signal dates is done as random circular shifts of the whole set of signal dates (10,000 shifts), which keeps the signals’ spacing and clustering; the p-value is the share of shifted sets whose mean forward return is at least the observed one.
3. **C3 data:** the warehouse holds daily FPI equity flows only from January 2021 (DII from April 2007), and no Nifty 500 market-capitalisation series. Weekly flows are therefore scaled by their own trailing 52-week standard deviation (known before the week), and the FPI tests cover 2021 onward only. A request for NSDL’s FPI history back to 1999 goes to lane 1.

**2026-10-14, before C10 was computed:**

1. **C10 entry:** the breadth reading and the index’s distance from its high are known only after a month-end’s close, so each forward return starts at the next session’s close (the audit E4 rule that C2 follows), not at the month-end close as first written.

**2026-10-22, before C50 was computed (data constraints found while wiring it):**

1. **C50 first cohort:** direct plans began on 1 January 2013, so the December 2015 cohort’s formation window runs from the end of January 2013 (35 months, annualised as such).
2. **C50 cohorts before 2018:** a fund that closed before SEBI’s 2018 recategorisation keeps its old label (“Growth”) in AMFI’s file and cannot be placed in a category, so it drops out of the December 2015 to 2017 cohorts entirely instead of counting as a failure. That flatters those cohorts’ survivors. The number of such funds alive at each formation date is reported, and the December 2018 and 2021 cohorts (after recategorisation) are reported separately from the December 2015 one.

**2026-10-24, before C60–C63 were computed (econ-models; the FCI as first built):**

1. **C60/C61 index design:** the FCI’s credit group first used the lending rate on new loans less the repo rate. Seen while building (index levels against known episodes only; no IIP, credit or GVA outcome had been computed): that spread narrows in every hiking cycle, because lending rates follow repo with a lag, and the index read the 2022-23 hikes as easing. The credit group now uses the level of the lending rate on new rupee loans. The index also needs at least three of its five groups in a month; the newest months, before the slower inputs arrive, have no value. Nothing else in C60–C63 changes.

**2026-10-24, before C71 was computed (a timing error found while wiring it):**

1. **C71 comparison print:** RBI fields the household survey in the first days of the survey month, before that month’s CPI release (around the 12th). The latest print a respondent could have seen is therefore the one for two months before the survey month, not one month as first written. The comparison uses that print; the one-month-back print is reported as a sensitivity.

## Results log

### 2026-10-02: C1, C2 and C3 (first runs, rules as specified and amended above)

Code: `pipeline/tipsheet/compute/evidence.py`; tables `evidence_c1_*`, `evidence_c2_*`, `evidence_c3_*` in `.cache/derived`.

**C1. Valuation and subsequent real returns (Nifty 500 TRI, CPI-deflated).**

| Signal | Horizon | Start months | Independent periods | Slope | R² | Newey-West t | Non-overlapping slopes (min / median / max; share negative) |
| --- | --- | --- | --- | --- | --- | --- | --- |
| Sensex CAPE10 (log) | 5 years | 2000-05 to 2021-08 | 5.3 | −31.7 | 0.69 | −5.4 | −46.5 / −28.5 / +18.7; 90% (4 points each) |
| Sensex CAPE10 (log) | 10 years | 2000-05 to 2016-08 | 2.6 | −11.3 | 0.76 | −25.1 | too few points (fewer than 3 per offset) |
| Nifty 500 CAPE10 (log) | 5 years | 2006-05 to 2021-08 | 4.1 | −14.2 | 0.35 | −3.4 | −51.3 / −6.6 / +112.5; 65% |
| Nifty 500 CAPE10 (log) | 10 years | 2006-05 to 2016-08 | 2.0 | −10.0 | 0.77 | −14.5 | too few points |
| Nifty 500 earnings yield | 5 years | 1999-02 to 2021-08 | 5.5 | +2.6 | 0.13 | +1.3 | −3.9 / +0.8 / +9.6; 40% |
| Nifty 500 yield gap | 5 years | 1999-02 to 2021-08 | 5.5 | +1.5 | 0.08 | +1.8 | −22.8 / +0.9 / +20.0; 33% |
| Nifty 500 yield gap | 10 years | 1999-02 to 2016-08 | 2.8 | −0.1 | 0.00 | −0.3 | too few points |

By Sensex CAPE tercile, the median real 5-year return was 12.8% a year from the cheapest third (CAPE 12 to 20), 7.9% from the middle and 2.3% from the dearest (CAPE above 24.4); each tercile holds about 1.4 independent 5-year periods. Tercile boundaries use the whole sample, so the table describes history and is not a trading rule.

- **Prediction C1 holds.** The CAPE slope is negative at both horizons. At 10 years the sample has 2.6 independent periods, so the t-statistic of −25 says nothing: overlapping windows make one or two episodes look like 196 observations. I do not call it significant.
- The 5-year CAPE result is the more informative one: nine in ten non-overlapping subsamples give a negative slope. Even there, a single valuation cycle (2007-08 peaks followed by weak returns, 2003 and 2009 troughs followed by strong ones) drives most of it.
- **The earnings yield and the yield gap carry little information** at either horizon: R² of 0.13 and 0.08 at 5 years, slopes that change sign across subsamples, and nothing at 10 years. Trailing earnings are noisier than ten-year average earnings, and the yield gap mixes in the bond yield, which moved for its own reasons.

**C2. Breadth thrusts (sample from 2003-09-11, when the breadth universe reached 200 stocks).**

Twelve signals: 2004-04, 2005-02, 2006-06, 2007-09, 2008-07, 2008-12, 2009-07, 2017-01, 2020-04, 2024-04, 2025-04 and 2026-04.

| Horizon | Signals | Mean after signal | Unconditional mean | Share positive after signal (unconditional) | p (circular-shift test) |
| --- | --- | --- | --- | --- | --- |
| 1 month | 12 | 2.7% | 1.4% | 75% (64%) | 0.25 |
| 3 months | 12 | 4.8% | 4.4% | 67% (65%) | 0.41 |
| 6 months | 11 | 15.5% | 8.8% | 82% (69%) | 0.09 |
| 12 months | 11 | 30.9% | 18.3% | 91% (78%) | 0.02 |

- **Prediction C2 holds:** fewer than 15 signals, a higher average 6-month return after a signal, and a p-value above 0.05 at 6 months.
- The 12-month result (p = 0.02) was not part of the prediction and leans on two signals that came straight after crashes (December 2008 and April 2020). With 11 events and four horizons tested, it is a lead worth tracking, not a finding. The July 2008 signal was followed by a 42% fall over three months.

**C3. Flows and returns (weekly; DII from October 2007, FPI from July 2021).**

| Flow | Sample | Flows predict next returns (F, p) | Returns predict next flows (F, p) | Sum of lagged-return coefficients |
| --- | --- | --- | --- | --- |
| DII | full | 1.1, 0.33 | 37.2, <0.001 | negative |
| DII | first half / second half | p = 0.36 / 0.29 | p < 0.001 in both | negative in both |
| FPI equity | full | 0.8, 0.36 | 44.7, <0.001 | positive |
| FPI equity | first half / second half | p = 0.29 / 0.62 | p < 0.001 in both | positive in both |

- **Prediction C3 holds.** Returns predict flows; flows do not predict returns. Foreign investors chase: the correlation of a week’s FPI flow with the previous week’s return is 0.47 and with the same week’s 0.33. Domestic institutions lean against the market: DII flows correlate −0.36 with the previous week’s return and −0.28 with the same week’s.
- Neither flow has a correlation above 0.13 in absolute value with any later week’s return (one to eight weeks ahead).
- Caveats: Granger tests measure predictability, not cause; the FPI sample is only five years; and the weekly sums mix trade dates and reporting dates.

### 2026-10-03: audit corrections (register `docs/research/AUDIT.md`, rows E1–E4, V3, V7)

- **C1, earnings-yield signal (V3, V7).** It now uses the consolidated-basis P/E, read from the maintained valuation table. Rescaling pre-2021 earnings by the switch factor scales the signal, so the slope falls and the t-statistic does not change:
  
  - 5 years: slope +2.64 → **+2.17**, t 1.27 (unchanged);
  - 10 years: +1.10 → **+0.92**.
  - The yield gap: 5-year slope +1.48 → +1.44 (R² 0.08 → 0.10).
  - CAPE rows are unchanged (they use IIMA’s series).
- **C1, inference (E1): superseded.**
  
  - **What C1’s inference got wrong:** Newey-West errors with 60 or 120 lags on 184 to 271 months are badly undersized. The persistent predictor’s small-sample (Stambaugh) bias was not measured. Adjacent non-overlapping offsets share 59 of 60 months, so “nine in ten subsamples negative” is not nine in ten pieces of evidence.
  - **What replaces it:** D1 (`audit_deepening_spec.md`) redoes the question properly. Its 5-year CAPE result has a bias-corrected bootstrap p of 0.055 (C1’s Newey-West t was −5.4), with about half the slope being small-sample bias, and an out-of-sample R² of +39% on 3.3 independent periods.
  - **Which statements stand and which do not:** C1’s prediction (a negative slope, not called significant at 10 years) still holds. Its sentence that “the 5-year CAPE result is the more informative one: nine in ten non-overlapping subsamples give a negative slope” overstates the evidence and is withdrawn.
- **C1, terciles and scatter (E2, E3).** The tercile table now carries 90% moving-block intervals (D2). Every tercile is flagged “too few” (under 5 independent periods). The scatter rows carry a `nonoverlap` flag marking one non-overlapping set.
- **C2, entry timing (E4).** Breadth is known only after the signal day’s close, so forward returns now start at the next close.
  
  - **Changes:** 1-month mean 2.7% → **2.0%** (p 0.25 → 0.38); 3-month 4.8% → **4.7%** (p 0.41 → 0.43); 6-month 15.5% → **14.9%** (p 0.09 → **0.11**); 12-month 30.9% → **30.7%** (p 0.02 → **0.03**).
  - **The 6-month mean now carries an interval:** 14.9% (90% interval 4.1–25.0%, 11 events) against 8.8% for all sessions.
  - **Prediction C2 holds.** The 12-month result stays exploratory.
- **C3:** unchanged.

### 2026-10-06: C9 (first run, rules as specified)

Data: NSDL’s 353 fortnightly reports, 31 January 2012 to 30 September 2026. A median of 24 sectors a fortnight pooled: 25 on NSDL’s own list, 19 on the BSE list. Placebo: 1,000 within-fortnight shuffles.

| Sample | Return | Fortnights | Mean rank corr. | t (NW 4) | Placebo p |
| --- | --- | --- | --- | --- | --- |
| Pooled | t−1 | 351 | +0.186 | 13.7 | <0.001 |
| Pooled | t | 353 | +0.235 | 19.8 | <0.001 |
| Pooled | t+1 | 351 | −0.007 | −0.6 | 0.52 |
| Pooled | t+2 | 349 | +0.015 | 1.2 | 0.22 |
| 2012–Apr 2022 | t−1 / t+1 / t+2 | 245 / 245 / 244 | +0.166 / −0.019 / +0.006 | 11.0 / −1.3 / 0.4 | <0.001 / 0.16 / 0.67 |
| From Apr 2022 | t−1 / t+1 / t+2 | 106 / 106 / 105 | +0.232 / +0.020 / +0.035 | 9.2 / 0.9 / 1.7 | <0.001 / 0.39 / 0.11 |

**Prediction C9 holds.** FPIs put money into the sectors that did well in the previous fortnight, in both eras. Their sector flows say nothing about the next fortnight’s sector returns, or the one after, which is the first an outsider could act on. The same-fortnight correlation is the largest. It is descriptive: flows and prices move together within a fortnight, and it says nothing about which drives which.

2026-10-07, wording only: the site is run by one person, so “we” and “our” now read “I” and “my”. No definition, test, threshold or result changed.

### 2026-10-14: C10 (first run, rules as specified and amended; laptop data to 2026-10-12)

**S1, narrow rallies** (Nifty 500 TRI within 5% of its 52-week high while the share of stocks above their 200-day average is below its expanding median): 28 of 290 month-ends from 2002-09, clustered (7 in 2019, 4 in 2025, 3 each in 2013, 2018 and 2026).

| Horizon | Flagged months | Mean after a narrow rally | Mean, other months | Median, flagged / other | Higher afterwards, flagged / other | Shift p (weaker) |
| --- | --- | --- | --- | --- | --- | --- |
| 3 months | 28 | −2.2% | +5.6% | −3.6% / +5.4% | 32% / 69% | 0.013 |
| 6 months | 27 | +0.5% | +10.9% | −1.7% / +8.4% | 44% / 75% | 0.035 |
| 12 months | 22 | +7.8% | +22.0% | +6.3% / +15.2% | 55% / 84% | 0.066 |

**S2, the equal-weight gap** (from 2006-03, 235 months, 19 independent years): slope of the forward 12-month return on the gap −0.29 (Newey-West t −1.66, R² 0.03), non-overlapping offsets −0.75 to +0.04. Median forward 12-month return by expanding tercile of the gap: low (the giants leading) 11.6% (3 independent years), middle 12.9% (4), high (the average stock leading) 7.9% (8).

**Grading:** the prediction held. Narrow-rally months had a lower mean 12-month return, and its shift p-value (0.066) is above 0.05; the S2 slope has |t| below 2. Not predicted: at 3 and 6 months the narrow-rally gap is larger and clears 0.05 (0.013 and 0.035). With three horizons tested on 28 clustered months, that is a lead to keep watching, not a finding; Bonferroni across the three horizons keeps only the 3-month result, narrowly. The equal-weight gap points the other way from the usual story, weakly: months after the average stock had been leading were followed by lower median returns.

### 2026-10-22: C11 (first run, rules as specified)

Sample 6 Oct 1999 to 14 Oct 2026. 10 washouts (9 with a full 12 months after), 16 euphoria signals. Code: `compute/evidence.py::c11_breadth_extremes`; bundle `evidence/breadth_extremes`.

|  | 3 months | 6 months | 12 months |
| --- | --- | --- | --- |
| After washouts, mean (share positive) | 9.1% (90%) | 12.2% (80%) | 15.7% (78%) |
| After euphoria, mean (share positive) | 7.9% (75%) | 10.5% (63%) | 16.4% (69%) |
| Any day, mean (share positive) | 4.2% (63%) | 8.3% (67%) | 17.8% (75%) |
| p, washout higher | 0.09 | 0.23 | 0.63 |
| p, euphoria lower | 0.89 | 0.74 | 0.46 |

**Graded:** fewer than 12 washouts, held (10). Fewer than 12 euphoria signals, failed (16). Washout 12-month mean above the unconditional mean, failed (15.7% against 17.8%). Washout p above 0.05 at 63 sessions, held (0.09). No reliable euphoria difference, held. Nothing clears the pre-set p < 0.01 bar. The three months after a washout were positive in 9 of 10 cases, against 63% of all three-month windows; with ten signals and six tests, that is a lead to watch, not a finding. Quintiles of the daily reading show no mean reversion: the highest fifth (above 81%) was followed by the highest median 12-month return (23%), each fifth holding about five independent years.

### 2026-10-22: C12 (first run, rules as specified and amended)

68 indices with at least three years both sides of launch (20 strategy, 22 thematic, 17 sectoral, 9 broad). Code: `compute/evidence.py::c12_post_launch`; bundle `evidence/post_launch`.

| Group | Indices | Excess over Nifty 500 before launch (mean, median) | After launch (mean, median) | Median decay | Share decayed | Sign-flip p |
| --- | --- | --- | --- | --- | --- | --- |
| Strategy | 20 | 3.6%, 4.5% | 1.1%, 0.2% | −3.1 pts | 80% | 0.01 |
| Sectoral | 17 | 4.4%, 3.9% | 1.1%, 0.3% | −2.0 pts | 71% | 0.14 |
| Thematic | 22 | 1.8%, 0.9% | 1.8%, 0.0% | −0.5 pts | 59% | 1.00 |
| Broad | 9 | 0.8%, 1.3% | 2.7%, 2.1% | +1.5 pts | 33% | 0.09 |

Pooled monthly excess return (Newey-West, 12 lags), pre-launch months against post-launch months: strategy 3.0% a year (t 1.9) against 0.1% (t 0.1); sectoral 6.5% (t 2.5) against −2.3% (t −0.9); thematic 3.6% (t 1.6) against −1.8% (t −1.1); broad 1.4% (t 0.5) against 1.4% (t 0.7).

**Graded:** strategy pre-launch excess above 2% a year, held (3.6%). Post-launch below half of it, held (1.1%). More than 60% decayed, held (80%). Smaller decay for sectors and themes, held. Broad indices within ±1 point, failed: they improved by 1.5 points, carried by mid and small caps after 2016. Caveats: most strategy launches fall in 2016–2020, so their post-launch years share one market era; sector decays are dominated by FMCG and Bank, whose short pre-launch windows (1996–99, 2000–03) were exceptional. Indices overlap in holdings and time, so the p-values are a guide.

### 2026-10-22: C20, C21 and C22 (first runs, rules as specified; data to 21 October 2026)

Code: `pipeline/tipsheet/compute/valuation_insights.py` (step `valuation_insights`); tables `evidence_c20_size_spread`, `evidence_c21_sectors`, `evidence_c21_months`, `evidence_c22_mcap_gdp` in `.cache/derived`; bundle `valuation/depth`.

**Implementation note, disclosed:** the first run filled gaps in the signal by carrying the previous month forward on both the signal and the return, which the D1 rule does not allow. The Nifty Smallcap 100’s P/B has two source gaps (June 2016 to January 2017, February to May 2020) and the Midcap 100 misses September 2020. The corrected rule (carry a lone one-month signal gap only; a longer gap, or any missing return, ends the sample) changed C20 from slope −18.10, bias-corrected −9.81, p 0.138, OOS R² −4.0% to the figures below. The verdict did not change. The Smallcap 100’s sample now ends in May 2016, so its 3- and 5-year rows have under two independent periods and are not shown on the page.

**C20. Midcap 100 relative P/B against its 3-year return over the Nifty 50** (2005-08 to 2023-09 starts, 218 months, 7.0 independent periods).

| Index | Horizon | Independent periods | Slope | Null mean slope | Bias-corrected | Bootstrap p | OOS R² |
| --- | --- | --- | --- | --- | --- | --- | --- |
| **Midcap 100 (hypothesis)** | **3 years** | **7.0** | **−18.5** | **−8.3** | **−10.2** | **0.128** | **−2.9%** |
| Midcap 100 | 1 year | 21.1 | −19.0 | −12.2 | −6.9 | 0.255 | −28.7% |
| Midcap 100 | 5 years | 4.2 | −6.5 | −6.7 | +0.1 | 0.509 | −12.7% |
| Smallcap 100 | 1 year (to 2015-05) | 5.1 | −42.6 | −37.0 | −5.6 | 0.437 | too few |
| Next 50 | 1 year | 23.7 | −19.3 | −4.7 | −14.6 | 0.119 | −61.9% |
| Next 50 | 3 years | 7.9 | −7.0 | −3.4 | −3.7 | 0.352 | −40.6% |
| Next 50 | 5 years | 4.7 | −3.5 | −2.8 | −0.7 | 0.475 | −13.7% |

**Grading: the prediction held.** The slope is negative (a relative P/B 10% higher went with mid caps trailing by about 1.8 points a year more over three years), and C20 does not pass: p 0.128, and the fitted line forecast worse than the historical-mean gap out of sample. About 45% of the in-sample slope is the bias a persistent signal produces with no real link.

**C21. Sectors cheap against their own P/B history minus dear, next 12 months** (2010-07 to 2025-09 starts, 183 months, 16.2 independent years, median nine sectors a month).

Mean spread +2.0 points (median +2.5), positive in 56% of start months; Newey-West t 0.61; placebo p 0.078 (2,000 within-month shuffles; the placebo’s 5th to 95th percentiles of the mean were −2.1 to +2.3). Reported, not tested: +5.3 points in the 66 months when dispersion was above its expanding median, +0.2 in the other 117.

**Grading: the prediction held.** The spread is positive and the placebo p is above 0.05; C21 does not pass.

**C22. Market value to GDP against the Nifty 50’s 5-year real return.**

| Signal | Start months | Independent periods | Slope | Null mean slope | Bias-corrected | Bootstrap p | In-sample R² | OOS R² (independent periods) |
| --- | --- | --- | --- | --- | --- | --- | --- | --- |
| (a) log market value to GDP | 1999-08 to 2021-09 | 5.4 | −14.7 | −9.4 | −5.3 | 0.225 | 0.64 | −37.2% (3.4) |
| (b) against its trailing 10-year average | 2009-04 to 2021-09 | 3.5 | −11.6 | −9.0 | −2.6 | 0.364 | 0.48 | +45.6% (1.5) |

**Grading:** the main prediction held: neither passes. Two sub-predictions failed. Signal (a) was not weak or wrong-signed in sample: its slope is negative with an R² of 0.64, though 64% of the slope is small-sample bias and it loses to the historical mean out of sample. Signal (b) has 3.5 independent periods, not fewer than three. Its positive OOS R² rests on 1.5 independent periods of forecasts and is not evidence.

### 2026-10-22: C30, C31 and C32 (first run; data to September 2026 for FPI, October 2026 for DII)

Code: `pipeline/tipsheet/compute/flows_scaled.py` (step `flows_scaled`); results `.cache/derived/flows_scaled.json`, published as `evidence/flows_scaled`. 10,000 permutations or shifts each.

**Deviations from the specification, all forced by the data and found while computing, disclosed here:**

- **Market cap starts in June 2009** (PRIME’s first quarter), so the scaled DII series starts in July 2009, not April 2007.
- **The DII daily series misses 33 sessions** since 2007 (9 in April 2007, the rest scattered, 24 of them in 2025-26: up to four in a month). A month missing up to five sessions is pro-rated by sessions traded ÷ sessions observed (12 months in the sample). Both C30 and C31 were rerun with those months left out; the conclusions do not change (figures below).
- **C32’s first flag is November 2015**, not December: the 36th month-end of the trailing twelve-month series is November 2015, and my “December” was an off-by-one in the text, not a different rule.
- The market-cap estimate matches the filings-based total within 0.5–2.0% at the five quarters where both exist.

**C30. Persistence (monthly flow in bp of market cap).**

| Flow | Sample | Months | Lag-1 autocorrelation | p | Same sign as last month | If independent |
| --- | --- | --- | --- | --- | --- | --- |
| FPI | 2012-01 to 2026-09 | 177 | 0.44 | <0.001 | 66.5% | 52.0% |
| FPI | first / second half | 88 / 89 | 0.43 / 0.37 | <0.001 / <0.001 | 70.1% / 62.5% | 55.8% / 50.2% |
| DII | 2009-07 to 2026-09 | 207 | 0.56 | <0.001 | 72.8% | 57.3% |
| DII | first / second half (split Jan/Feb 2018) | 103 / 104 | 0.44 / 0.56 | <0.001 / <0.001 | 62.7% / 82.5% | 50.4% / 72.7% |

Without the pro-rated months: DII 0.52 full, 0.41 first half, 0.47 second half. Domestic institutions have been net buyers for 38 consecutive months (August 2023 to September 2026), against a previous longest run of 22 (to January 2019). The longest FPI selling run was 9 months (to June 2022).

**Grading:** the prediction held. FPI autocorrelation is above 0.2 with p below 0.05, and sign persistence beats independence; DII flows are more persistent than FPI flows, and more so in the second half.

**C31. Same-month coincidence (177 months, 2012-01 to 2026-09).** Correlation of scaled FPI flow with the Nifty 500 TRI return 0.52 (p < 0.001); DII −0.54; FPI with DII −0.86. FPI-selling months (71): mean return −1.07%, down 63% of the time; buying months (106): +2.79%, down 20%; gap −3.86 points (p < 0.001). Among selling months, DII buying at least as large as the FPI selling in 56: mean −1.14%; smaller in 15: −0.81%; gap −0.34 points (p = 0.81). Without pro-rated months the gap is +0.62 (p = 0.67).

**Grading:** the prediction held. The same-month correlation is above 0.3 with p below 0.01, selling months average lower returns, and absorption does not mark better months. Descriptive only: it does not say which moved which.

**C32. What followed heavy foreign selling (month-ends 2015-11 to 2025-09, 119 months, 9.9 independent years).** Flagged 54 months (45%, not a fifth: the scaled twelve-month flow drifted down relative to its own expanding history, so “bottom fifth so far” fired often). Mean forward 12-month return 15.2% after flagged months, 15.9% after the rest (median 10.9% / 10.5%; share higher 83% / 86%); gap −0.8 points, circular-shift p = 0.98. Slope +0.29 points of return per 1% of market cap of trailing flow, Newey-West t 0.11; non-overlapping offset slopes from −9.3 to +12.9. Flagged runs: Nov 2015–Jul 2016 (forward 21.2%), Nov–Dec 2016 (34.6%), Aug–Oct 2017 (3.9%), Mar 2018–Mar 2019 (2.3%), Mar–May 2020 (67.3%), Jan 2022–Mar 2023 (18.6%), Jan–Sep 2025 (4.0%). September 2026 is flagged.

**Grading:** the null held (p above 0.05, |t| below 2); the direction I expected did not: flagged months were marginally worse, not better. The flag’s high firing rate is a weakness of the rule as written, recorded here rather than changed after the fact. On this record heavy foreign selling has not been a buying signal or a warning.

### 2026-10-22: C40 to C43 and C4 (first runs, rules as specified; bhavcopy to 2026-10-14)

Code: `pipeline/tipsheet/compute/ipo_size.py` (step `page_ipo_size`); checks in `.cache/derived/ipo_size_checks.json`.

**Size validation passed before any return was computed.** For all 117 listings of 2025–26 with an XBRL share count filed within 45 days, the AMFI-implied market value at listing was within 10% of the filed count times the listing close (median ratio 1.000; the worst, Shadowfax, 1.004). 435 of 539 mainboard listings could be sized; 410 from 2018-01-16 got peers.

**A data fix found on the way (not a change of rule).** The price panel starts a new canon at a rename that kept the ISIN. Burger King (BURGERKING → RBA, February 2022, same ISIN) therefore looked as if it had stopped trading 14 months after listing, and every IPO table exited it at its last close. `compute/ipo.py` now stitches such continuations (one mainboard and two SME listings were affected). The H1 alpha moved from +4.18% to +4.24% a year (t 0.91 → 0.92); the 3-year mean BHAR against Nifty 500 from +17.6% to +17.3%.

**C40, IPOs against size-matched peers.**

| Measure | Value |
| --- | --- |
| Calendar-time months (2018-05 to 2026-10), median IPOs a month | 102, 96 |
| Mean monthly IPO minus peers | +0.17 points (annualised +2.5%) |
| Newey-West t (3 lags), p | 0.48, 0.63 |
| Months the IPO portfolio led | 52% |

| Buy-and-hold, from the listing close | 1 year (n = 360) | 3 years (n = 180) |
| --- | --- | --- |
| Median IPO return / median peer-basket return | +0.3% / +8.6% | +20.4% / +89.8% |
| Mean BHAR against peers | +0.4 points | −18.9 points |
| Median BHAR against peers | −10.5 points | −52.6 points |
| Share beating their peers | 42% | 33% |
| Median BHAR against Nifty 500 (same issues) | −8.5 points | −22.7 points |

- **Prediction C40 holds.** In calendar time IPOs are indistinguishable from seasoned companies of the same size. The typical IPO trails its peers, more so than it trails Nifty 500, and only a third beat their peers over three years.
- The 3-year buy-and-hold figures rest on the 2018–2023 cohorts, and 132 of the 180 are from 2021–2023, whose three years covered the mid and small-cap boom: their peers’ median return was +114% (2021), +95% (2022) and +64% (2023).
- **Exploratory, not pre-registered (the split was chosen after seeing the yearly figures):** the calendar-time difference was +1.6 points a month in 2018–2020 (32 months, a median of 30 IPOs, t 2.3) and −0.5 points a month from 2021 (70 months, t −2.0). The full-sample null averages a few strong early months, when the portfolio was small, with five years of the IPO portfolio trailing its peers (2021 to 2025 each). A lead to keep watching, not a finding.
- The H2 flipping result survives size-matching: Spearman between the listing gain and the 1-year BHAR against peers is 0.000 (bootstrap 95% interval −0.12 to 0.12).

**C41, big and small IPOs.** Spearman between log size and the 1-year BHAR against peers: −0.014, t −0.26, p 0.79 (n = 360 in 74 listing months; bootstrap interval −0.13 to 0.10). **Prediction C41 fails on direction** (I expected small IPOs to do worse) and the result is null either way. By size tercile (cuts at ₹2,119 crore and ₹5,981 crore):

| Tercile | Median size | Median listing gain | Median 1-year return from issue | At or above issue after 1 year | Median 1-year BHAR vs peers | Beat peers |
| --- | --- | --- | --- | --- | --- | --- |
| Small | ₹1,009 cr | +7.3% | +6.1% | 53% | −5.5 points | 42% |
| Middle | ₹3,508 cr | +17.4% | +32.6% | 63% | −12.7 points | 41% |
| Large | ₹12,792 cr | +16.9% | +29.8% | 68% | −12.9 points | 43% |

Small IPOs did worse for applicants, with smaller listing gains and a weaker year from the issue price. Against companies of their own size they did no worse than large IPOs: every tercile trailed its peers at the median, and the means are close to zero in all three.

**C42, the promoter’s stake.** Spearman between the promoter stake at the first filing and the 1-year BHAR against Nifty 500: +0.076, t 1.31, p 0.19 (n = 297 listings from 2021, 49 listing months; bootstrap interval −0.03 to 0.17). Against size-matched peers: +0.073, t 1.26. **Prediction C42 holds:** positive and not significant. The eleven listings with no promoter at all (Zomato, CarTrade, PB Fintech, Paytm, Delhivery, Tamilnad Mercantile Bank, SAMHI, IndiGene, ixigo, FirstCry, Swiggy) had a median 1-year return from the issue price of −19% and a median BHAR against Nifty 500 of −45 points, against −2 to −16 points for the other groups. Eleven issues, eight of them listed in 2021 and 2024, so it is a description of one cohort, not a rule.

**C43, survivorship.** After the rename fix, 4 of 539 mainboard listings stopped trading on NSE: Equitas Holdings (2023) and Ujjivan Financial Services (2024), both merged into their small-finance banks; TCNS Clothing (2024), merged into Aditya Birla Fashion; and Future Supply Chain Solutions (2025), which was already down 99.6% from its listing close when it stopped. None stopped inside a 1-to-36-month or 3-year window, so treating them as total losses changes nothing: H1 alpha +4.24% a year either way, 3-year mean BHAR +17.3% either way. **Prediction C43 holds.** The universe excludes 26 mainboard rows of NSE’s issue list: 17 without an issue price, 6 whose symbol never traded within 15 sessions, 2 not yet listed, 1 BSE-first listing. Survivorship is not what drives the IPO results here; the caveat that matters is that NSE’s list starts in 2016, so no IPO in the sample has yet lived through a long bear market.

**C4, the IPO cycle and later market returns** (signal: trailing 12-month IPO money as a share of NSE market value, used from the end of the following month).

| Horizon | Start months | Independent periods | Slope (return per point of signal) | R² | Newey-West t | Non-overlapping slopes (min / median / max; share negative) |
| --- | --- | --- | --- | --- | --- | --- |
| 12 months | 2013-03 to 2025-08 (150) | 12.5 | −0.37 | 0.09 | −1.86 | −0.64 / −0.36 / −0.13; 100% |
| 36 months | 2013-03 to 2023-08 (126) | 3.5 | −0.13 | 0.01 | −0.71 | −3.47 / −0.36 / +1.05; 72% |

By expanding tercile of the signal, the median forward 12-month Nifty 500 TRI return was +17.9% (low, 1.1 independent years), +20.9% (middle, 3.1) and +8.8% (high, 5.3). **Prediction C4 holds:** negative at 36 months and not significant. The 12-month slope is negative in every one of the 12 non-overlapping offsets but its t is −1.86 on 12.5 independent years: busy IPO years have been followed by below-average years more often than not, without the evidence to call it a signal.

**Grading summary.** C40 holds, C41 fails on direction (null), C42 holds (null), C43 holds, C4 holds (null). None of the three primary tests clears 0.0167.

### 2026-10-22: C50 to C53 (first runs, rules as specified and amended; Data bank snapshot of 2026-10-22, NAVs to 2026-10-21, TRIs to 2026-10-14)

Code: `pipeline/tipsheet/compute/fund_scorecard.py` (C50, C51) and `fund_flows_lab.py` (C52, C53); tables `fund_c50_*`, `fund_c51_*`, `fund_c52_*`, `fund_c53_*` and the checks files `fund_scorecard_checks.json`, `fund_flows_lab_checks.json` in `.cache/derived`. Sample: 673 active equity schemes with paired direct and regular growth plans, 87 of them closed or merged since 2013.

**Data fixes between the first run and this one** (all general rules, recorded in `docs/methods/fund_scorecard.md`; no rule of the tests changed):

1. The plan pairing dropped any scheme whose name was shared by a long-dead code (UTI Large Cap and UTI Value among them); codes that stopped before 2013 are now ignored and the latest-alive code wins. Seven schemes came back.
2. **Re-coded schemes.** When a fund house changed hands, AMFI sometimes closed the old plan codes and opened new ones for the same scheme (L&T to HSBC on 28 November 2022; IDBI to LIC, Principal to Sundaram, BNP Paribas to Baroda BNP Paribas, JPMorgan to Edelweiss). The first run counted the old codes as closed funds and the new ones as fresh launches. Codes that printed the same ISIN, one stopping at most 31 days before the other starts, are now one fund, the NAV history joined at the hand-over. 27 active schemes were re-coded; 29 false closures and 21 false launches went away.
3. An IDCW plan spelled out in full (“Income Distribution cum Capital Withdrawal”) of a scheme with “Growth” in its name passed as a growth plan; such plans are now excluded.

Effect: the five-year share of diversified direct plans beating their index rose from 63% to 67% and five-year survival from 85% to 92%; three-year figures did not move. No grade below changed between runs.

**C50, persistence.** Non-overlapping cohorts (December 2015, 2018, 2021; 726 fund-cohorts, 185 in the top quarter):

| Cohorts | Top quarter again | Top quarter beat benchmark next | Bottom quarter beat next | Top / bottom survived | Mean rank corr. | Shuffle p |
| --- | --- | --- | --- | --- | --- | --- |
| 2015, 2018, 2021 (headline) | 21.1% | 45.4% | 54.7% | 98.9% / 95.3% | −0.012 | 0.60 |
| 2018, 2021 (after recategorisation) | 19.8% | 49.6% | 62.5% | 98.5% / 94.1% | −0.057 | 0.86 |
| All eight, overlapping | 19.9% | 45.1% | 55.4% | 99.0% / 96.3% | −0.057 | – |

Mean rank correlation by cohort: 2015 +0.08, 2016 −0.02, 2017 −0.23, 2018 −0.07, 2019 −0.15, 2020 −0.14, 2021 −0.04, 2022 +0.12. Funds alive but uncategorisable (pre-2018 label) at formation: 32, 18, 15, 10, 10, then none.

**Grading: the prediction held** on all three counts (top quarter again below 35%, mean correlation inside −0.1 to +0.2, shuffle p above 0.05). Not predicted: the bottom quarter beat its benchmark next more often than the top quarter (55% against 45%), a hint of reversal that the rank correlation does not confirm.

**C51, new active equity funds** (launches February 2019 to September 2025; cumulative excess return; 95% intervals resample launch quarters):

| Horizon | Group | Funds | Median lead over Nifty 500 TRI | Interval | Beat it |
| --- | --- | --- | --- | --- | --- |
| 2 years (headline) | All | 207 | +7.7 pts | +5.3 to +11.1 | 73% |
| 2 years | Diversified | 103 | +7.9 | +4.9 to +11.5 | 77% |
| 2 years | Sectoral and thematic | 104 | +7.5 | +3.3 to +12.1 | 69% |
| 1 year | All | 268 | +4.1 | +2.4 to +5.7 | 67% |
| 3 years | All | 146 | +12.9 | +6.9 to +20.2 | 72% |

**Grading: the prediction failed.** I predicted the median new fund would lag the Nifty 500 over two years, fewer than half would beat it, and thematic launches would lag diversified ones by more. The median new fund led by 7.7 points and 73% beat it; thematic launches did about as well as diversified ones. Exploratory checks added after seeing this (not pre-registered): against each diversified fund’s own category index the two-year lead is +5.2 (+3.2 to +7.8); against the median older fund in the same category over the same month-ends, all new funds led by +4.5 (+2.6 to +6.9) at two years and +4.1 (+0.2 to +9.4) at three. So part of the lead is the era and the category mix (mid and small caps led for much of 2020 to 2024), and part is new funds beating older ones in their own categories, which at three years is only just clear of zero. Every launch met a mostly rising market.

**C52, NFOs and the market** (signal months August 2019 to August 2025, 69 with a forward return, about 5.8 independent years; pause months May and June 2022 dropped, along with every window touching them):

- Correlation of NFO intensity with the trailing 12-month Nifty 500 return: +0.56 (81 months).
- Slope of the next 12 months’ return on intensity: −15.5 (R² 0.13), Newey-West t −2.16; across the 12 non-overlapping offsets the slope runs from −40.8 to +26.9.
- Expanding top third of intensity (17 months: July 2021 to April 2022, October 2022, February 2024 and June to October 2024): mean forward return 3.8% against 18.5% for other months.

**Grading: half held.** Intensity rose with past returns as predicted (correlation above +0.3). The forward slope is negative as predicted, but its t (−2.16) is beyond the |t| < 2 I predicted. I do not read that as a finding: the offset slopes change sign, and the top-third months come almost entirely from two launch waves.

**C53, return-chasing across nine equity categories:**

| Return window | Months | Mean rank corr. | NW t (12) | Placebo p | Positive months |
| --- | --- | --- | --- | --- | --- |
| Previous 12 months | 76 | +0.34 | 4.70 | <0.001 | 79% |
| Same month | 88 | +0.16 | 3.67 | <0.001 | 65% |
| Next 12 months, from t+2 | 75 | +0.10 | 0.91 | 0.003 | 53% |

**Grading: the prediction held.** Flows chase the past year’s category returns (t above 2) and do not predict the next year’s on the registered statistic (|t| below 2). The placebo p for the forward test is low because it shuffles labels month by month and ignores the overlap of consecutive twelve-month windows; the Newey-West t is the registered test and accounts for it.

### 2026-10-24: C60–C63 (econ-models; rules as pre-registered and amended in item 7)

Code: `pipeline/tipsheet/compute/economy_validation.py` (run on demand); results in `.cache/derived/economy_validation.json`. Data as refreshed on 2026-10-24 (credit, macro, pulse and rbi steps rerun first). Seed 20261024; 10,000 bootstrap draws.

One data constraint, found while running: the held GVA series (`macro_gdp_quarterly`, bases linked) switches to the 2022-23 base from Q1 FY24, so “old base alone” runs to Q4 FY23, not FY25 as written.

**C60. FCI tightening and later IIP growth: does not pass. Prediction holds (negative slope, no pass).**

| Horizon | FCI slope | Newey-West t | R² (with control) | Months | Independent periods | Non-overlapping slopes (min / median / max; share negative) |
| --- | --- | --- | --- | --- | --- | --- |
| 3 months | −0.66 | −0.87 | 0.22 | 113 | 37.7 | −1.29 / −0.99 / +0.45; 67% |
| **6 months (the test)** | **−0.31** | **−0.39** | **0.14** | 107 | 17.8 | −2.60 / −1.01 / +2.49; 67% |
| 12 months | −2.10 | −1.37 | 0.29 | 95 | 7.9 | −10.31 / −4.15 / +4.12; 64% |

Sample May 2014 to February 2026 (h = 6), lockdown windows dropped. Benchmark (10-year less 1-year RBI month-end spread, same regression): slope +0.83, t +0.97, R² 0.01 over its longer sample from 2005. Over the FCI’s own sample, R² 0.137, the same as the FCI’s 0.137. Nearly all the fit comes from the control (IIP growth’s own recent change), so the FCI adds nothing to it at six months. After months when the FCI tightened by more than one standard deviation (17 of them), IIP growth changed by −0.08 pp over the next six months, against −0.22 pp after other months. The sign is negative at every horizon and in two-thirds of the non-overlapping subsamples, but nothing is distinguishable from zero.

**C61. FCI tightening and later bank credit growth: does not pass. Prediction wrong** (I expected a negative slope with t near −2).

| Horizon | Slope | t | R² | Months |
| --- | --- | --- | --- | --- |
| 3 months | +0.36 | +0.81 | 0.09 | 116 |
| **6 months (the test)** | **+0.09** | **+0.09** | 0.13 | 110 |
| 12 months | −0.72 | −0.45 | 0.01 | 98 |

There is no sign that a tightening of financial conditions, as measured, precedes slower bank lending within a year.

**C62. The activity composite against GVA: does not pass. Prediction holds** (positive, above 0.4, not better than the best single indicator).

- All quarters from Q1 FY14 to Q1 FY27 outside FY21–22 (45): the composite’s correlation with real GVA growth is 0.46. IIP alone gives 0.72, then core industries 0.64 and car registrations 0.63. Composite less IIP: the 95% interval is −0.48 to −0.06, so the composite is reliably **worse** than IIP, not merely no better.
- Old base alone (32 quarters to Q4 FY23): composite 0.52, IIP 0.82; interval of the difference −0.43 to −0.07.
- Reading: equal weights give the payment, travel and credit themes as much say as industry, and those move with their own trends more than with GVA. The composite answers “are the indicators running hot against their own past”, which is a different question from GDP growth. It should not be presented as a GDP tracker.

**C63. Breadth and GVA acceleration: prediction wrong. Breadth is neither coincident nor leading.**

|  | Correlation | 95% interval | Quarters |
| --- | --- | --- | --- |
| Lag 0 (breadth at quarter-end against that quarter’s acceleration) | +0.08 | −0.03 to +0.30 | 46 |
| Lag 1 (against the next quarter’s) | −0.12 | −0.57 to +0.20 | 43 |
| Old base only, lag 0 | +0.07 | −0.04 to +0.33 | 34 |

The share of indicators improving says little about whether quarterly GVA growth sped up or slowed. GVA growth, revised and on two bases, may be too noisy a target, and breadth measures direction in three-month averages rather than year-on-year acceleration. Either way, the front door should present breadth as a description of the indicators, not as a read on GDP.

### 2026-10-24: C70 and C71 (first runs, rules as specified and amended; MoSPI CPI to September 2026, RBI household survey to September 2026)

Code: `pipeline/tipsheet/compute/economy_prices.py` (`c70`, `c71`); results in `.cache/derived/economy_prices.json` under `tests`.

**C70, forecasting headline twelve months ahead** (origins January 2017 to September 2025, 105 months, about nine independent years; Diebold-Mariano against core, positive t means worse than core):

| Forecast | RMSE (pts) | MAE | Mean error | DM t vs core | p |
| --- | --- | --- | --- | --- | --- |
| Headline now (random walk) | 2.34 | 1.90 | −0.02 | +2.86 | 0.004 |
| **Core ex food and fuel** | **1.61** | **1.28** | −0.16 |  |  |
| Weighted median, 23 groups | 1.95 | 1.55 | +0.43 | +1.13 | 0.26 |
| Trimmed mean, 10% each tail | 1.97 | 1.60 | +0.07 | +1.61 | 0.11 |
| Trimmed mean, 25% each tail | 1.89 | 1.51 | +0.25 | +1.05 | 0.29 |
| Ecumenical median (long-history readings) | 1.93 | 1.52 | +0.10 | +1.56 | 0.12 |

Without origins from April 2019 to March 2021 (81 origins): core 1.54, headline 2.29 (t +2.26), median 1.67 (t +0.46), trimmed 25% 1.68 (t +0.54), trimmed 10% 1.83 (t +1.18), ecumenical 1.77 (t +1.30).

**Grading: mostly held, one part failed.** Core beat the headline random walk, decisively. No underlying measure beat core; every one did worse, by margins the test cannot tell from chance. The trimmed means did no better than core. **The ecumenical median’s RMSE was 20% above core’s (15% without the Covid origins), outside the 10% I predicted.** A likely reason, not tested: its readings include headline at three horizons, which carries food’s noise into the median; core alone, being an exclusion measure, drops it. This fits the RBI’s finding (George et al., 2024) that exclusion measures held up best in India. The test covers only the coarse base-2012 versions (23 groups); the 358-item measures from January 2026 have no targets yet.

**C71, households’ one-year-ahead expectation** (73 rounds, March 2012 to September 2025, targets to September 2026; about five rounds a year):

- Mean expectation minus realised inflation: **+5.37 points**, Newey-West t 11.6.
- RMSE of the expectation: 5.83 points (MAE 5.37); of the latest print respondents could have seen: 2.27 (MAE 1.88). Diebold-Mariano t +5.0, p below 0.001. The one-month-back print (sensitivity): RMSE 2.31.

**Grading: held.** Households overpredict by more than five points on average and their median expectation is a much worse forecast of next year’s inflation than the latest print. Read the survey as a measure of how households feel about prices, and of its changes, not as a forecast.

### 2026-10-24: C65 and C66’s backtest (econ-models; pseudo-real-time, latest vintage)

Code: `pipeline/tipsheet/compute/economy_nowcast.py`; rows in bundle `economy/growth/nowcast_backtest`. Real GVA growth, RMSE in pp; Diebold–Mariano p on squared errors, h = 1.

|  | Quarters | Model, 3 months | Model, 2 months | Random walk | AR(1) | Mean of last 4 | DM p, model (3m) vs random walk |
| --- | --- | --- | --- | --- | --- | --- | --- |
| Bridge (C64), old base | 29 | 1.63 | 1.70 | 1.57 | **1.42** | 1.65 | 0.90 |
| Bridge (C64), new base | 9 | **0.66** | 0.66 | 0.79 | 0.79 | 0.72 | 0.21 |
| Challenger (C66), old base | 29 | 2.26 | 2.43 | 1.57 | **1.42** | 1.65 | 0.02 (worse) |
| Challenger (C66), new base | 9 | 0.82 | 0.67 | 0.79 | 0.79 | **0.72** | 0.89 |

Old-base quarters: Q1 FY17–Q4 FY20 and Q2 FY23–Q2 FY26; new base: Q1 FY25–Q1 FY27. GDP (3 months): bridge 1.88 against the random walk’s 1.84 on the old base; 0.58 against 0.66 on the new base.

- **C65 does not pass. Prediction wrong** (I expected the bridge to beat the random walk modestly on the old base). On the old base the bridge is worse than the random walk and the AR(1). It over-predicted every quarter from Q1 FY18 to Q4 FY19, by 1.3 to 3.1 pp: through the 2017–19 slowdown, old-base GVA grew well below what IIP, cement, steel, transport and credit implied. On the new base it is the best of the four over 9 quarters, but the p of 0.21 cannot separate that from luck.
- **C66’s backtest prediction is wrong.** The no-fit challenger is worse than the bridge on the old base and significantly worse than the random walk (p 0.02): with the slope fixed at 1, it carries indicator swings into GVA one for one. On the new base it ties the random walk.
- **C66’s Q2 FY27 prediction is wrong.** The challenger’s first nowcast (GVA 9.7%) is above the bridge’s (9.1%), not below. With nothing fitted, the indicators alone still point to growth above 9%; the bridge’s slopes are not what puts it above the RBI’s 7.2%.
- Both models stay published and scored, labelled experimental, as C64 and C66 set out.

### 2026-10-24: C80–C82 (first run, rules as specified; RBI lending and deposit rates to August 2026, repo moves to October 2026)

Five complete cycles have WALR fresh at baseline: easing 2015-17 (−200 bp, Jan 2015 to Aug 2017), tightening 2018 (+50 bp), easing 2019-20 (−250 bp), tightening 2022-23 (+250 bp), easing 2025 (−125 bp, Feb to Dec 2025). The October 2026 rise (+25 bp) opens a sixth.

**C80, external benchmarks: prediction wrong.** WALR-fresh pass-through six months after the last move: easing 2015-17 95%, tightening 2018 blank (the next cycle began first), easing 2019-20 68%; tightening 2022-23 79%, easing 2025 64%. The 2015-17 easing beats both later cycles. Two things the rule did not anticipate: the 2015-17 easing ran 31 months from first move to last, so most of its pass-through had happened before the six-month clock started; and in the 2025 easing new-loan rates had passed through 84% by the month of the last cut and then rose again (to 58% by August 2026), while credit growth rose from about 10% to 18%.

**C81, margins: half held, half wrong.** Tightening 2022-23: new deposits 92%, new loans 79% six months after the last move (held). Easing 2025, at August 2026: new deposits 76%, new loans 58% (wrong: deposit rates fell further than loan rates, so the fresh-business margin widened in the easing instead of narrowing).

**C82: pending.** Baseline September 2026 (not yet published); graded on the April 2027 figures.

**Reported after grading, not tests:** months by which each rate trailed the repo in reaching half the cycle’s move (the rate’s months to half less the repo’s own): new loans 8, 1 and 6 months in the three earlier cycles; 3 and 0 in the two later ones. This is the measure on which external benchmarks look faster; it was chosen after seeing C80, and with five cycles it is a description.

### 2026-10-24: C85, C86 and C87 (econ-households; rules as pre-registered; RBI surveys to September 2026, PLFS to Apr–Jun 2026 quarterly, EPFO to July 2025, ESIC to July 2026, NACH to September 2026)

Code: `pipeline/tipsheet/compute/economy_households.py` (C85) and `economy_jobs.py` (C86, C87); results in `.cache/derived/economy_households.json` and `economy_jobs.json`, and bundles `economy/households/c85`, `economy/jobs/c86`, `economy/jobs/c87`.

**C85, urban confidence and household consumption.** The workbook and India Data Hub agree on the CSI in all 67 overlapping rounds (largest gap 0.004), so the test ran. 39 usable quarters, Q2 FY14 to Q1 FY27; 23 evaluated out of sample from Q2 FY18.

|  | RMSE, pp |
| --- | --- |
| Random walk | 1.67 |
| AR(1) | 1.69 |
| AR(1) plus the four-quarter change in CSI | 1.72 |

Diebold–Mariano, model against the AR(1): stat +0.37, p 0.71. In-sample Spearman of the CSI change with PFCE growth 0.14 (of the CSI level, 0.46).

**Grading: does not pass, as predicted; one part of the prediction wrong.** Confidence adds nothing out of sample: the model is 1.8% worse than the AR(1) (within the 5% I predicted) and no better than the random walk. But the in-sample correlation of the change, 0.14, is below the 0.2 to 0.5 I predicted: urban sentiment barely moves with consumption growth even in hindsight. The level does better (0.46), which fits sentiment trending with the cycle rather than marking its turns.

**C86, households’ feelings about jobs and the unemployment rate.** 13 four-quarter changes, Q3 FY20 to Q1 FY27, after the lockdown quarters are dropped.

- Spearman of the changes: **+0.10**, permutation p 0.75. Without the changes spanning the January 2025 redesign (10 changes): +0.31, p 0.38.
- Spearman of the levels (22 quarters): −0.24.

**Grading: does not pass, as predicted.** Changes in how urban households feel about jobs do not move with changes in measured urban unemployment; the sign is even the wrong way. The levels lean the expected way, but weakly (−0.24): unemployment fell from 9.7% in Apr–Jun 2018 to 6.7%, while households’ net view of jobs stayed negative. Thirteen changes are few; this says the two do not track, not that they are unrelated.

**C87, NACH bulk credits and formal hiring.** Three-month sums, growth on a year earlier.

| Payroll measure | Months | r with NACH other credits | Circular-shift p |
| --- | --- | --- | --- |
| EPFO net new subscribers | 40 (Apr 2022 to Jul 2025) | −0.36 | 0.59 |
| ESIC new insured workers | 44 (Apr 2022 to Jul 2026; the eight touching ESIC’s missing February and March 2024 drop out) | −0.24 | 0.27 |

**Grading: does not pass, as predicted.** NACH other credits do not track net new EPF or ESIC subscribers; if anything they move the other way, though not by more than chance allows. The pages label NACH credits a flow of salaries, pensions, dividends and interest, never a count of jobs.

### C95. Does fuller manufacturing capacity precede faster fixed investment? (2026-10-24, economy/business)

**Registered before calculating any forward investment outcome.** Seen beforehand: source schemas, survey series names and date coverage, and the latest two national-accounts rows while checking the input contract. No capacity/investment lead comparison has been computed. C96–C99 remain unused.

- **Question:** does seasonally adjusted OBICUS capacity utilisation add information about later real GFCF growth beyond the investment growth already observable at a conservative decision date?
- **Data:** held RBI OBICUS workbooks, seasonally adjusted utilisation; held economy Growth quarterly GFCF growth, separate GDP bases. These are latest-vintage revised histories, not historical first releases. The current survey snapshot has only one vintage. This is a pseudo-real-time timing exercise, never a live forecast record or causal claim.
- **Rule and timing:** capacity quarter q is assumed available at quarter-end +120 days (conservative assumed lag, not an archived release date). At that decision date the latest completed GFCF quarter allowed is q+1 (assumed release at quarter-end +60 days). The outcome is GFCF’s year-on-year real growth in q+3, the first complete calendar quarter starting after that decision date. Require complete calendar matches, never row shifts across missing quarters. Use only completed outcomes as of the run date. Exclude a row if any quarter from q through q+3 lies between April 2020 and June 2022. No interpolation or seasonal adjustment of our own.
- **Test:** OLS outcome = intercept + capacity level + GFCF growth in q+1. Fit independently on each GDP base, never pool or splice bases. At least 24 complete origins per base. Report capacity slope (percentage points of growth per percentage point of utilisation), sample dates/count and a 95% percentile interval from 1,999 moving-block bootstrap draws, four consecutive calendar quarters per block, seed 95. Missing calendar quarters break eligible bootstrap blocks. A positive slope with its entire interval above zero passes; otherwise does not pass. Too few rows is untestable. No alternative horizon, threshold or model selected after seeing results.
- **Prediction:** fuller capacity will have a positive slope on the old base, but its interval will include zero once existing investment growth is accounted for. The new base will be too short to judge. All results and grades will be published equally prominently.

**C95 timing correction, 2026-10-24, after the first computation and before publication:** the +120-day decision date above cannot observe q+1 GDP with its assumed +60-day release lag (a quarter plus that lag is about 150 days). The initial implementation failed to check this inequality. The first run produced an old-base slope of −1.2017 (interval −2.2345 to −0.0953; 40 origins) and 11 new-base origins; those were seen before this correction. Set the decision delay to **165 days**, explicitly require the control’s quarter-end +60 days to be no later than the decision, and require the target to start after the decision. This keeps q+3 as the first full subsequent quarter and leaves the regression, outcome horizon, exclusions, bootstrap and grading unchanged. Also require target quarter-end +60 days by the run date, a stricter completion requirement. This is a disclosed timing repair, not a second model selected for its result; report the corrected result and whether it changed.

### 2026-10-24: C95, business capacity and later investment (first run, with disclosed timing repair)

Code: `pipeline/tipsheet/compute/economy_business.py`; results in `economy/business/facts`, `tests.c95`. Latest-vintage, assumed release delays, calendar-matched origins; **not a real-time forecasting record**. The timing correction disclosed above changes no usable pair or numerical result.

- **2011-12 base: does not pass; prediction wrong.** 40 origins, Q4 FY12–Q3 FY25. Capacity coefficient −1.2017 percentage points of later real investment growth per point of utilisation, conditional on q+1 investment growth; 95% moving-block interval −2.2345 to −0.0953. The prediction was positive but indistinguishable from zero. Its sign was wrong; the registered positive-association test fails.
- **2022-23 base: untestable; prediction held.** 11 origins, Q4 FY23–Q2 FY26. Below the registered minimum of 24, so no slope or interval is presented. This is not confirmation of either a positive or negative relationship.

The negative old-base association is descriptive and conditional; no alternative lag, model or trading interpretation has been selected after seeing it. Both results are displayed on the capacity page.

## C110–C111. States: do the monthly state readings track state output, and does a lead last? (`/economy/states/`, econ-indicators)

Written 2026-10-26 by econ-indicators, which takes C110–C119. Code will be `pipeline/tipsheet/compute/economy_states.py`.

- **What I had seen beforehand, disclosed in full:** an exploratory pass on 26 October, before this block, computed for the 19 table states (at least 1% of all states’ nominal GSDP) the per-year Spearman correlation of full-financial-year growth in GST, e-way bills and vehicle registrations with nominal and real GSDP growth (GST FY20 to FY25: 0.43, 0.71, 0.52, 0.40, −0.15, 0.04 against nominal; e-way FY21 to FY25: 0.29, 0.47, 0.00, 0.01, −0.14; vehicles FY06 to FY25 between −0.54 and 0.58), and the median across months of the Spearman correlation of states’ three-month growth with the same reading twelve months earlier (GST −0.03, e-way 0.22, power 0.40 on 15 months, vehicles 0.00, UPI 0.20 on 16 months). No permutation test, pooled statistic or pandemic exclusion had been computed. Because these were seen, the predictions below are written from them and are not a test of foresight; what the blocks fix in advance is the sample, the statistic and the pass line, so the page cannot pick a flattering cut.

### C110. Do a state’s monthly readings move with its output?

- **Data:** GSTN collections by state (CGST, SGST, IGST and cess), GSTN e-way bills (within and out of the state) and VAHAN registrations by state, as read by `economy_states`; RBI Handbook nominal GSDP (DBIE 1575), newest base each state has, growth never computed across a base change.
- **Measure:** each reading summed over a complete financial year (all twelve months held), growth on the year before, against the state’s nominal GSDP growth in the same year.
- **Sample:** the table’s states; every financial year where both exist for at least 15 of them, leaving out FY21 and FY22 (lockdown and its base effects). Electricity (from 2023) and UPI (from April 2023) have at most one usable year and are reported, not graded.
- **Test:** for each reading, the mean across years of the cross-state Spearman correlation. p: 10,000 permutations shuffling state labels within each year independently, one-sided (share of permuted means at least the observed).
- **Pass:** mean ρ at or above 0.4 with p below 0.05.
- **Prediction C110:** fails for all three. GST comes closest (mean ρ about 0.2 to 0.3, after dropping its strongest year, FY21); e-way bills and vehicles stay near zero.

### C111. Does a state’s lead on a monthly reading last a year?

- **Data:** the five monthly readings of the States table (GST, e-way bills, electricity demand, vehicle registrations, UPI), each as the latest three months against the same three months a year earlier, as on the page.
- **Measure:** for each month t with the reading for at least four in five table states, the cross-state Spearman correlation of growth at t with growth at t − 12 (the two windows share no month).
- **Sample:** months t whose two windows and their year-earlier bases, the 27 months from t − 26 to t, avoid April 2020 to March 2022 entirely (t from June 2024 on, or t by March 2020).
- **Test:** the mean ρ across months. p: 10,000 permutations of state labels applied once to all t-side values (keeping each state’s own time path, so neighbouring months stay correlated as in the data), one-sided.
- **Pass:** mean ρ at or above 0.3 with p below 0.05. Electricity and UPI have fewer than 24 months and are reported, not graded.
- **Prediction C111:** fails for GST, e-way bills and vehicles (mean ρ below 0.25): which state is ahead on a monthly reading one year says little about which is ahead the next. UPI, a payment habit that spreads state by state, may persist, but is too short to grade.

### 2026-10-26: C110 and C111 (rules as specified above; run 26 October 2026, 19 table states, 10,000 permutations, seeds 110 and 111)

Code: `pipeline/tipsheet/compute/economy_states.py` (`track_output`, `lead_persistence`); tables `economy_states_signals_years`, `economy_states_signals_months` in `.cache/derived`.

**C110, do the monthly readings move with state output?**

| Reading | Years | Mean ρ with nominal GSDP growth | p | Graded |
| --- | --- | --- | --- | --- |
| GST collected | 4 (FY20, FY23 to FY25) | 0.18 | 0.062 | yes |
| E-way bills | 3 (FY23 to FY25) | −0.04 | 0.62 | yes |
| Vehicle registrations | 17 (FY06 to FY25, less FY12, FY21, FY22) | 0.16 | 0.003 | yes |
| UPI payments | 1 (FY25) | 0.07 | 0.39 | no |

**Grading: does not pass for any reading, as predicted.** Vehicles’ link is more than chance but far below the pass line, and it swings from −0.54 (FY25) to 0.58 (FY07). GST’s mean fell below the predicted 0.2 to 0.3 because FY19 had no complete prior year (GST began in July 2017), so its years are fewer than the exploratory pass used.

**C111, does a state’s lead last a year?**

| Reading | Months | Mean ρ, t against t − 12 | Share of months above zero | p | Graded |
| --- | --- | --- | --- | --- | --- |
| GST collected | 34 (Sep 2019 to Mar 2020, Jun 2024 to Aug 2026) | 0.13 | 62% | 0.049 | yes |
| E-way bills | 27 (Jun 2024 to Aug 2026) | 0.15 | 63% | 0.11 | yes |
| Vehicle registrations | 209 (Mar 2005 to Sep 2026) | 0.02 | 53% | 0.32 | yes |
| Electricity demand | 12 | −0.27 | 17% | 1.00 | no |
| UPI payments | 16 | 0.20 | 63% | 0.098 | no |

**Grading: does not pass for any graded reading, as predicted** (all mean ρ below 0.25). GST’s p is just under 0.05 but its mean is less than half the pass line. Shown on `/economy/states/signals/`; the overview’s count of readings above the median now says it is a snapshot and links there.

## C120–C121. F&O expiry: does the index settle near “max pain”? (`/fo/premium/`, fo-insights)

Written 2026-10-26 by fo-insights, which takes C120–C129, **before anything below was computed**. Nothing about max pain, strike-level open interest or expiry settlements had been looked at in this or any earlier session. Code will be `pipeline/tipsheet/compute/fo_expiry.py`.

- **The belief tested:** Indian trading forums hold that option writers “pull” the index towards the max-pain strike into expiry, so max pain the day before tells you where the index settles.
- **Data:** NSE F&O bhavcopy, contract level (`nse_fo_bhavcopy*`, January 2020 on, lot sizes as in `fo_analytics`); index options on NIFTY (primary) and BANKNIFTY (secondary). Underlying close from the bhavcopy (UDiFF) or Nifty index history (legacy layout), as in `fo_analytics`.
- **Max pain on session t for expiry E:** using only contracts of expiry E open at the close of t, the listed strike K\* that minimises the total value option holders would receive if the index settled at K\*: Σ calls OI × max(K\* − K, 0) + Σ puts OI × max(K − K\*, 0), OI in units. Ties: the strike nearest the underlying close.
- **Outcome:** S\_E, the underlying close on expiry day (NSE’s final settlement price for index options).
- **Sample:** every expiry E of the index, weekly and monthly, with t the last trading session before E (k = 1). Secondary: t five sessions before E (k = 5), skipped where t would fall on or before the previous expiry of the same index. Expiries with fewer than 10 listed strikes holding OI at t are dropped and counted.
- **C120 (pull):** regress the move to expiry, (S\_E − S\_t) / S\_t, on the gap to max pain, (K\*\_t − S\_t) / S\_t, across expiries, with no intercept restriction. One observation per expiry; consecutive weekly expiries do not overlap at k = 1. Heteroskedasticity-robust (HC1) standard errors. **Pass:** slope above 0 with one-sided p below 0.05 for NIFTY at k = 1, and the slope at k = 5 of the same sign.
- **C121 (direction):** share of expiries where the index moved towards max pain (sign of S\_E − S\_t equals sign of K\*\_t − S\_t; expiries with K\*\_t within 0.05% of S\_t left out), against 50%, exact binomial, one-sided. **Pass:** above 50% with p below 0.05.
- **Benchmarks reported beside it:** median absolute distance of S\_E from K\*\_t and from S\_t (a random walk’s forecast), as % of S\_t.
- **Prediction C120–C121:** both fail for NIFTY. Max pain sits close to the spot price on most days, so it “predicts” the settlement about as well as the spot does, which is why the belief survives; the slope is small and not significant, and the direction share is within 47–55%. BANKNIFTY the same.
- **Not a test:** the share of index-option contracts open the day before expiry that expire worthless (no intrinsic value at S\_E), by month, is descriptive and published without a pass line.

### 2026-10-26: C120 and C121 (rules as specified above; run 26 October 2026, `compute/fo_expiry.py`)

- **NIFTY, k = 1:** 355 expiries, 2020-01-02 to 2026-10-19 (1 dropped for fewer than 10 strikes with OI). Slope of the move to expiry on the gap to max pain 0.098 (HC1 se 0.089, t 1.10, one-sided p 0.135), R² 0.011. Moved towards max pain in 140 of 318 (44.0%, one-sided p 0.986). Median distance of the settlement from max pain 0.63% of the index, from the previous close 0.52%.
- **NIFTY, k = 5:** only 4 expiries qualify, because with weekly expiries the session five before almost always falls on or before the previous expiry. Slope 0.67. The secondary condition is uninformative on this sample; this was not foreseen when the block was written and the rule is left as written.
- **BANKNIFTY, k = 1:** 277 expiries; slope 0.105 (p 0.34); towards max pain 49.2% (p 0.62). k = 5: 27 expiries, slope -0.43.
- **Grades:** C120 **fails** (slope not significant). C121 **fails** (towards in 44%, below half). Both fail as predicted, though the direction share fell below the predicted 47–55% range. Max pain is a worse guess of the settlement than the previous close.
- **Sensitivity, not graded:** without the two expiries whose gap to max pain exceeded 3% (19 and 26 March 2020, when books opened before the crash put max pain 3.9% and 20.2% above the index), the NIFTY slope is −0.439 (HC1 t −2.00, n 353): if anything the index moved away from max pain. Neither this nor the direction share below 50% was a pre-registered test, and neither is claimed as a finding.
- **Descriptive:** 85.5% of index-option contracts open the session before expiry expired worthless (2020-01-02 to 2026-10-19); by premium value that session, 24.6%.

## C122–C123. “Sell in May and go away”: do November–April returns beat May–October in India? (`/indices/`, insights-nine)

Written 2026-10-27 by insights-nine, which takes C122–C129. **Disclosure:** part of this has already been computed. A survey on 26 October 2026 ran the C122 calculation below on the Nifty 500 TRI, 1995–2025, and saw the result: November–April beat May–October in 16 of 31 years, mean excess return 4.8% against 4.5% a half-year, sign-flip p 0.96. C122 is graded on that seen result and labelled as such. C123 had not been computed or looked at, on any index, when this block was written. Code will be `pipeline/tipsheet/compute/season.py`.

- **The belief tested:** the market does better from November to April than from May to October, so investors should sell in May and buy back in November.
- **Data:** month-end NSE index total-return levels (`index_tri`), and the 91-day T-bill total-return index (`asset_tri_daily`) for cash.
- **Halves:** for year y, May–October is the six monthly returns May y to October y, and November–April is November y to April y+1. A half is kept only when all six months exist. Each half’s excess return is the compounded index return less the compounded T-bill return over the same six months.
- **Statistic:** for each year with both halves, d\_y = excess(Nov–Apr) − excess(May–Oct). Test the mean of d\_y against zero by sign-flip permutation (20,000 draws, seed 20261027), two-sided. Report the count of years with d\_y > 0.
- **C122 (seen, Nifty 500, from 1995):** **Pass** (the effect exists) if mean d\_y > 0 with p below 0.05. Expected: fail, as already seen.
- **C123 (unseen):** the same test on the Nifty Midcap 150 TRI and Nifty Smallcap 250 TRI (from 2005) and the Nifty 50 TRI (from 1999). **Pass** if any one of the three has mean d\_y > 0 with p below 0.05/3.
- **Prediction C123:** all three fail. Half-year means are noisy: with 20–27 years, a 3–4 point gap a half-year is well inside chance.
- **Reported, not graded:** each half’s average excess return, the share of half-years with positive excess, and the same numbers before and after 2010.

### 2026-10-27: C122 and C123 (rules as specified above; run 27 October 2026, `compute/season.py`)

- **C122, Nifty 500 (seen before registration):** 31 years, 1995 to 2025. November–April beat May–October in 16. Mean excess return 4.9% a half-year for November–April against 4.5% for May–October; mean difference 0.37 points, sign-flip p 0.94. **Fails**, as already seen. Through 2010 November–April led (8.4% against 3.7%, 16 years); from 2011 May–October led (5.4% against 1.1%, 15 years).
- **C123 (unseen):** Nifty Midcap 150, 21 years from 2005: November–April beat May–October in 8, mean difference −3.86 points, p 0.54. Nifty Smallcap 250, 21 years: 9 of 21, −5.58 points, p 0.37. Nifty 50, 26 years from 2000 (the TRI starts June 1999, so 1999’s May–October half is incomplete): 11 of 26, −3.09 points, p 0.54. None comes near the bar of p below 0.017, and all three point the other way. **Fails**, as predicted.
- **Reading:** there is no sell-in-May effect on NSE indices. In every index from 2011, May–October has been the better half on average, which is not significant either and is not claimed.

## C130–C134. The F&O ban list: does a stock’s entry into the ban period follow its move, or lead the next one? (`/fo/`, nse-crowding)

Written 2026-10-29 by nse-crowding, which takes C130–C139, **before any return, volatility or ban-entry outcome below was computed**. What had been looked at: single days of `combineoi` (2010, 2011, 2012, 2014, 2018, 2021, 2024, 28 Oct 2026) and `fo_secban` to learn the format, and that the ban flag in day t’s `combineoi` matches the `fo_secban` list for t+1 (28/29 Oct 2026). Nothing about prices around bans. Code will be `pipeline/tipsheet/compute/fo_crowding.py`.

- **Data:** `nse_fo_mwpl` (NSE `combineoi_DDMMYYYY.zip`, 2010-02-01 on): for every F&O stock and session, the market-wide position limit (MWPL), open interest across exchanges and the next session’s limit, “No Fresh Positions” while the stock is in the ban period. Utilisation u = open interest / MWPL. Prices: Tipsheet’s split- and bonus-adjusted bhavcopy panel (`equity_prices_adjusted`, close × adj; price only, no dividends), matched on (date, NSE symbol).
- **Amended 2026-10-29, before any outcome was computed:** the Data bank tests showed that recent files print the next-day limit as 95% of MWPL less the *future-equivalent* (delta-based) open interest, not the notional one. So u uses the open interest the ban rule used on each date: future-equivalent wherever the file prints it, notional before. Nothing else changes.
- **Universe:** U(t) = stocks in day t’s `combineoi` with a close on t. Membership comes from each day’s file, so stocks later dropped from F&O or delisted are kept while they were in it.
- **Event (ban entry):** stock s at session t with the ban flag in t’s file (the ban applies from t+1) and no flag in its previous file. **Fresh entries** (primary sample) also had no flag in the 20 sessions before t. The file is posted about 19:00 IST on t, so the first close an investor could trade at is t+1.
- **Excess return:** the stock’s return minus the equal-weighted mean return of U(t) over the same window. A stock with no close at the window’s end is dropped from that test and counted.
- **C130 (does the ban follow a move?):** pre-event excess return over the 20 sessions to t (close t−20 to close t). **Pass** (bans follow unusual moves) if the median absolute pre-event excess return of fresh entries is more than twice the median absolute 20-session excess return of all stock-days in U(t) over the same years. Reported beside it: the share of entries after a rise, and the change in utilisation over those 20 sessions.
- **C131 (does it lead the return?):** forward excess return from close t+1 to close t+21 (20 sessions), fresh entries. Mean across events; sign-flip permutation clustered by calendar month of t (all events of a month flip together), 20,000 draws, seed 20261029, two-sided. **Pass** (a return signal) if p < 0.05. Secondary, not graded: 5 and 60 sessions.
- **C132 (reversal of the move into the ban?):** across fresh entries, OLS slope of the C131 forward excess return on the C130 pre-event excess return. 95% interval from a month-block bootstrap (resample calendar months with replacement, 10,000 draws, seed 20261029). **Pass** if the interval excludes zero.
- **C133 (does volatility fall after?):** per fresh entry, log of the standard deviation of daily log returns over t+2..t+21 over that of t−19..t, less the median of the same log ratio across U(t) for the same windows. Mean across events; sign-flip permutation clustered by month as in C131. **Pass** if p < 0.05.
- **C134 (do concentrated client positions lead to a ban?):** only if NSE’s `mwpl_cli_DDMMYYYY.xls` (clients holding at least 3% of a stock’s MWPL, from 2012) is available over the sample. Stock-days 2012 on with u(t) in \[0.60, 0.95) and no ban flag at t. Treated: at least one client listed for the stock on t. Outcome: a ban flag in any of the next 20 sessions’ files (t+1..t+20). Strata: 5-point bands of u × calendar year. Statistic: the difference in outcome rate, treated minus untreated, averaged across strata with both groups present, weighted by n\_treated·n\_untreated/(n\_treated+n\_untreated). 95% interval from a month-block bootstrap (2,000 draws, seed 20261029). **Pass** if the difference is above zero and the interval excludes zero. **Decision rule:** the client file becomes a Data bank dataset and a page reading only if C134 passes; otherwise the result is logged here and on the page as a null, and the file is not collected.
- **Predictions:** C130 passes: bans come after large moves, mostly rallies with open interest building. C131 fails: the mean forward excess return is within ±1.5% and not significant. C132: the slope is negative (some reversal), but I do not expect it to clear the bar. C133 passes, negative: entries come at volatility peaks and volatility falls back relative to peers. C134: the difference is positive and small (a few points); whether it passes I cannot say.
- **Descriptive, no pass line:** stocks in the ban period per session since 2010; the share of F&O stocks above 80% utilisation; ban lengths; the stocks nearest the 95% line on the latest session.

### 2026-10-29: C130–C134 (rules as specified and amended above; `combineoi` 2010-02-01 to 2026-10-28, `mwpl_cli` 2012-01-02 to 2026-10-28, `compute/fo_crowding.py`)

- **Sample:** 4,139 sessions; 2,179 ban entries, 739 of them fresh, in 179 stocks. Ban spells: median 4 sessions, 90th percentile 10. Every fresh entry has a close at t.
- **Implementation note, disclosed:** twelve `combineoi` files (28 May to 13 June 2012) are corrupt at NSE. The first run counted a stock flagged on 14 June 2012 as a fresh entry against its last file before the gap. The corrected rule, written after that first run’s outcomes had been seen, needs a stock’s previous file to be at most seven days older. It removed three entries (742 fresh to 739, 2,182 in all to 2,179). No verdict changed.
- **C130:** median absolute pre-event excess return 9.05% against 4.88% for all stock-days in the same years, a ratio of 1.85. **Fails** (the bar is 2). Predicted to pass. 49% of entries came after a rise (median +9.1%) and 51% after a fall (median −9.0%), against the predicted “mostly rallies”. Median utilisation 0.70 twenty sessions before entry.
- **C131:** mean forward excess return +0.62% over 20 sessions (median −0.57%, 47% positive), 738 events in 174 months, p 0.21. **Fails**, as predicted. One entry has no close at t+21 and is dropped.
- **C132:** slope +0.013, 95% interval −0.071 to +0.105. **Fails**, as predicted, but the predicted negative sign did not appear.
- **C133:** mean log volatility ratio relative to peers −0.091 (median −0.085; 41% positive), p 0.00005; volatility after entry is about 8% lower relative to before than for other F&O stocks. **Passes**, as predicted. The pre-window ends at the move that triggered the ban, so this is partly that burst fading; it is not evidence that the ban itself calms a stock.
- **Secondary, not graded:** 5 sessions +0.78% (p 0.009); 60 sessions +0.63% (p 0.51). The 5-session reading is one of two secondary windows and is not claimed.
- **C134:** 57,929 stock-days in 177 months and 105 strata, from 3,630 sessions with a client file; 50,876 (88%) treated. A ban within 20 sessions followed 32.6% of treated stock-days and 23.3% of untreated. Stratified difference +4.74 points, 95% interval 1.24 to 8.17. **Passes**. Predicted positive and small. A research run on the same day, on files parsed before three source quirks were found (a client-count column on 1 January 2013, misnumbered headers, a stale tail on 28 June 2018), gave +4.66 (1.21 to 8.10); the verdict and decision did not depend on the repair. Under the decision rule, `mwpl_cli` is now the Data bank dataset `nse_fo_mwpl_clients` and the board shows each stock’s listed clients. The comparison group is small (12% of stock-days), and the result is an association: a client near the 3% line is a sign of crowding, not shown to cause a ban.
- **Reading:** bans come after big moves in either direction and are followed by ordinary returns. Nothing here supports trading on the ban list.

This is `docs/research/site_calculations_spec.md`. The specification was committed before any result was computed; changes after that are logged in it with dates and reasons.

---

Historical research, not investment advice. This page in HTML: https://tipsheet.markets/research/site_calculations_spec/. Every published dataset: https://tipsheet.markets/data/v1/manifest.json. A map of the site: https://tipsheet.markets/llms.txt
