Trend, momentum and risk models, second review: pre-registered specification

Written 2026-10-13, before any result below was computed; the commit that adds this file is the freeze. Changes after results are logged at the bottom with the date and the reason. Brief: docs/briefs/trend_tab_and_models.md, stage 2. Builds on trend_momentum_spec.md (five single rules, none beat buy-and-hold after deflation), trend_barometer_spec.md (the 24-vote ensemble, TR-H1 to TR-H5), momentum_barometer_spec.md (MO-H5 sector rotation, failed) and market_mood_spec.md (HAR baseline for Nifty 500 volatility).

What the literature supports, and what it does not

Every reference below was fetched and its title, authors, year and the cited claim checked on 2026-10-13 (abstract or publisher record; where only the abstract was seen, the claim cited is the abstract’s).

Volatility is forecastable; this is the most robust result in the list.

• Bollerslev, “Generalized Autoregressive Conditional Heteroskedasticity”, Journal of Econometrics 31(3), 1986: the GARCH model.
• J.P. Morgan/Reuters, RiskMetrics Technical Document, 4th ed., 1996: the exponentially weighted moving average with decay 0.94 for daily data.
• Hansen and Lunde, “A Forecast Comparison of Volatility Models: Does Anything Beat a GARCH(1,1)?”, Journal of Applied Econometrics 20(7), 2005: of 330 ARCH-type models, none clearly beat GARCH(1,1) for exchange rates; models with a leverage effect did better for a stock.
• Corsi, “A Simple Approximate Long-Memory Model of Realized Volatility”, Journal of Financial Econometrics 7(2), 2009: the HAR model, a sum of daily, weekly and monthly volatility components, forecasts well and has become the benchmark.
• Patton, “Volatility Forecast Comparison Using Imperfect Volatility Proxies”, Journal of Econometrics 160(1), 2011: rankings are robust to a noisy proxy only under certain losses (MSE and QLIKE); QLIKE has more power.

Trend and momentum: real in large diversified futures portfolios, much weaker market by market, and partly a volatility-scaling effect.

• Moskowitz, Ooi and Pedersen, “Time Series Momentum”, Journal of Financial Economics 104, 2012: 12-month own-return momentum in 58 futures; a diversified portfolio earns abnormal returns and does best in extreme markets.
• Hurst, Ooi and Pedersen, “A Century of Evidence on Trend-Following Investing”, Journal of Portfolio Management 44(1), 2017: 1-, 3- and 12-month trend across 67 markets since 1880, volatility-targeted.
• Levine and Pedersen, “Which Trend Is Your Friend?”, Financial Analysts Journal 72(3), 2016: time-series momentum and moving-average crossovers are near-equivalent linear filters.
• Kim, Tse and Wald, “Time Series Momentum and Volatility Scaling”, Journal of Financial Markets 30, 2016: the 2012 result comes mainly from volatility scaling; without it, time-series momentum is close to buy-and-hold.
• Huang, Li, Wang and Zhou, “Time Series Momentum: Is It There?”, Journal of Financial Economics 135(3), 2020: asset-by-asset regressions show little predictability in or out of sample; the strategy’s profit is matched by one built on historical means, which needs no predictability.
• Baltas and Kosowski, “Demystifying Time-Series Momentum Strategies: Volatility Estimators, Trading Rules and Pairwise Correlations” (SSRN 2140091; in Market Momentum: Theory and Practice, Wiley, 2020): better volatility estimates and continuous signals cut turnover by more than a third without a significant cost to performance.
• Goulding, Harvey and Mazzoleni, “Momentum Turning Points”, Journal of Financial Economics 149(3), 2023: slow and fast momentum define bull, correction, bear and rebound states; blending the two speeds earns positive alpha.
• George and Hwang, “The 52-Week High and Momentum Investing”, Journal of Finance 59, 2004: nearness to the 52-week high forecasts stock returns better than past returns, without long-run reversal.

Volatility management of equity factors: in-sample alpha, out-of-sample doubts.

• Moreira and Muir, “Volatility-Managed Portfolios”, Journal of Finance 72(4), 2017: scaling down when volatility is high raises Sharpe ratios for the market and several factors.
• Barroso and Santa-Clara, “Momentum Has Its Moments”, Journal of Financial Economics 116(1), 2015: scaling stock momentum by its own forecast volatility nearly eliminates crashes and nearly doubles its Sharpe ratio.
• Daniel and Moskowitz, “Momentum Crashes”, Journal of Financial Economics 122(2), 2016: momentum crashes follow market declines in high volatility and are partly predictable.
• Cederburg, O’Doherty, Wang and Yan, “On the Performance of Volatility-Managed Portfolios”, Journal of Financial Economics 138(1), 2020: across 103 equity strategies, volatility-managed portfolios do not beat their unmanaged versions out of sample, despite positive spanning alphas.

Persistence, regimes and testing.

• Lo and MacKinlay, “Stock Market Prices Do Not Follow Random Walks: Evidence from a Simple Specification Test”, Review of Financial Studies 1(1), 1988: the variance-ratio test with a heteroskedasticity-robust statistic.
• Lo, “Long-Term Memory in Stock Market Prices”, Econometrica 59(5), 1991: once short-range dependence is allowed for, no evidence of long memory; the classical rescaled-range (Hurst) estimate is biased.
• Hamilton, “A New Approach to the Economic Analysis of Nonstationary Time Series and the Business Cycle”, Econometrica 57(2), 1989: Markov-switching regimes.
• White, “A Reality Check for Data Snooping”, Econometrica 68(5), 2000; Hansen, “A Test for Superior Predictive Ability”, Journal of Business & Economic Statistics 23, 2005: testing the best of many models against a benchmark; SPA has more power.
• Bailey and López de Prado, “The Deflated Sharpe Ratio”, Journal of Portfolio Management 40(5), 2014.
• Politis and Romano, “The Stationary Bootstrap”, Journal of the American Statistical Association 89(428), 1994.

Indian evidence. I found no peer-reviewed test of time-series momentum on NSE indices. Cross-sectional stock momentum is documented in India (Sehgal and Jain on BSE 500 stocks, 2011 and 2015; Chui, Ranganathan, Rohit and Veeraraghavan, “Momentum, Reversals and Liquidity: Indian Evidence”, Pacific-Basin Finance Journal, 2023, intermediate and long-term momentum on the BSE 2000–2021). Those are stock-level, long-short results and do not carry over to timing an index. Our own earlier tests on NSE indices are the main Indian evidence, and they are negative.

Rejected without a test, and why.

• Downside beta (Ang, Chen and Xing, “Downside Risk”, Review of Financial Studies 19, 2006): a cross-sectional premium among stocks, not a way to time or describe an index; a replication (Atılgan, Demirtaş and Günaydın, European Financial Management 26(2), 2020) finds it does not survive value-weighting or controls.
• Hurst-exponent persistence: Lo (1991). Variance ratios (TM-H4) are used instead.
• Dynamic crash-weighted momentum (Daniel and Moskowitz): built for long-short stock momentum with forecasts of its mean; the products here are long-only index funds. Its volatility half is tested in TM-H1.
• Cross-sectional sector momentum: already tested and failed (MO-H5).
• Correlation-adjusted trend (Baltas and Kosowski): designed for long-short futures portfolios. Their signal-construction result (continuous signals) is already how the 24-vote ensemble trades.
• New signal types or horizons for the trend score: Levine and Pedersen show the common filters are near-equivalent; the registered score stays.

Data and periods (all hypotheses)

• Markets: the 128 markets of the trend section (compute/trend_section.load_markets): live NSE total-return indices, gold in rupees, the NSE 5-year G-sec index, the rupee and 11 overseas indices in rupees. Strategy tests use the 116 investable markets (NSE indices, gold, G-secs). Forecast and persistence tests use all 128.
• Returns: daily log returns of the level for volatility; excess returns over the 91-day T-bill cash index for trend and momentum, as in the existing score.
• Back-calculated and live history. For each NSE index, sessions before its launch date (trend_launch_dates.csv) are back-calculated. Every result is reported post-launch only (the primary sample; indices with an unknown launch date are left out of it) and on the full history. “Live” means after this freeze: the chosen models are logged from 2026-10-13 and their live record is reported from the first full quarter.
• Out of sample. Every estimated model is refit each January on all data to the previous 31 December (expanding window) and used for that calendar year. The first out-of-sample year for a market is the first January with at least 1,260 sessions (five years) of its history.
• No look-ahead. Forecasts and signals use the close of day t. Strategies trade at the close of t+1 (trend.backtest). Costs as in the existing replays: 0.10% each way; fund costs equity 0.15%, gold 0.50%, G-sec 0.20%, cash 0.20% a year.
• Inference. Stationary bootstrap (Politis and Romano), mean block 21 sessions for daily series and 4 weeks for weekly, 2,000 resamples, resampling calendar dates jointly across markets so cross-market correlation is kept. “Both halves” splits each market’s out-of-sample period at its middle date. Seed 20261013.
• Multiple testing. Holm across the seven hypotheses’ primary p-values at family α = 0.05. For strategy hypotheses, the deflated Sharpe ratio counts every trial in this file plus the 84 of trend_barometer_spec.md: N = 84 + 18 (TM-H1) + 2 (TM-H3) + 3 (TM-H5) = 107. Forecast hypotheses use Hansen’s SPA over all their variants.

Risk

RK-H1: a volatility forecast beats trailing volatility

• Target: realised variance over the next 21 sessions, the sum of squared daily log returns from t+1 to t+21, forecast at the close of t.
• Benchmark: trailing 63-session realised variance scaled to 21 sessions. (The site’s “volatility, past year” uses 252 sessions; it is reported too, as a weaker benchmark.)
• Candidates: (1) EWMA, decay 0.94, no estimation; (2) GARCH(1,1), Gaussian quasi-likelihood with variance targeting, refit each January, the 21-session forecast from its term structure; (3) HAR, 21-session-ahead average variance regressed on the average squared return over the last 1, 5 and 22 sessions (ordinary least squares, refit each January, forecasts floored at a tenth of the trailing 252-session variance).
• Loss: QLIKE, L = RV/F − log(RV/F) − 1, on every session of the out-of-sample period.
• Pass, for each candidate, all of:
  1. the median across markets of (benchmark loss − candidate loss) is above zero, with its 95% bootstrap interval above zero;
  2. lower loss than the benchmark on at least 75% of markets;
  3. the same in both halves (1 and 2 each half);
  4. the same in the post-launch-only sample;
  5. neighbourhood: EWMA decay 0.92, 0.94, 0.96 and 0.97, and HAR monthly window 22 and 63, all pass (1);
  6. Hansen’s SPA on the cross-market mean loss differential, all seven variants against the benchmark, p < 0.05.
• Choice for the site: among passing candidates, the lowest pooled loss; if intervals overlap, the simplest (EWMA, then HAR, then GARCH).

RK-H2: a calibrated range for the next month

• From the chosen RK-H1 forecast σ (21-session), the 90% range of the next 21-session log return is ±1.645σ (RK-H2a, normal), or the 5th and 95th percentiles of past standardised 21-session returns, r/σ, from the expanding window (RK-H2b, filtered historical).
• Coverage on non-overlapping 21-session blocks of the out-of-sample period.
• Pass: pooled coverage between 87% and 93%, Kupiec’s unconditional-coverage test not rejected at 5% on at least 80% of markets, and both halves inside 85% to 95%. If both pass, RK-H2a (simpler) is used. Expected shortfall (the mean of returns beyond the 5% bound) is reported only for a passing variant.

Trend and momentum

TM-H1: volatility scaling improves the trend rules (and is it the scaling or the trend?)

• Strategies, long or cash per market: exposure = s × min(1, σ* / σ̂), where σ̂ is the annualised EWMA (0.94) volatility at t and σ* the expanding median of the market’s own σ̂ (no look-ahead).
  • VS-ensemble: s = trend score / 100.
  • VS-12m: s = 1 when the 12-month excess return is positive, else 0.
  • VS-hold: s = 1 (volatility-managed buying and holding, as in Moreira and Muir, long-only).
• Each is compared with its unscaled counterpart (the ensemble, the 12-month rule, buying and holding) and with buying and holding.
• Measure: Sharpe ratio of daily returns over cash, after costs.
• Pass, for each strategy, all of:
  1. the median across investable post-launch markets of Sharpe(strategy) − Sharpe(buying and holding) above zero, its 95% bootstrap interval above zero;
  2. the deflated Sharpe ratio of the equal-weight cross-market series of (strategy − buying and holding) above 0.95, N = 107;
  3. both halves;
  4. neighbourhood: σ* × 0.75, 1 and 1.25, and EWMA decay 0.94 and 0.97 (six variants), all with a positive median in (1).
• Attribution (reported, not a pass condition): if VS-hold passes as well as VS-ensemble or VS-12m, the gain is from the scaling, not the trend (Kim, Tse and Wald).
• After tax (the lab engine) is reported beside every result. A strategy that passes before tax but trails buying and holding after tax for a 30% taxpayer is published as that, not as a signal.

TM-H2: nearness to the 52-week high ranks indices

• Universe: the 23 sector indices (least overlapping); broad and strategy indices reported as a second table.
• Each Friday, rank the live, post-launch members by close ÷ highest close of the last 252 sessions. Hold the top third and the bottom third, equal weight, for 4 weeks (four overlapping tranches).
• Pass: mean weekly top-minus-bottom return above zero with Newey-West p < 0.05 (lags 4) and the 95% bootstrap interval above zero; both halves; neighbourhood (high window 126 or 252 sessions, hold 4 or 13 weeks) all positive.

TM-H3: blending slow and fast trend (turning points)

• Per market: slow = 12-month excess return positive; fast = 1-month excess return positive. Exposure = 1 when both are up (bull), 0 when both are down (bear), 0.5 when they disagree (correction or rebound). This is the 50/50 blend of the two speeds.
• Pass: as TM-H1 (1)–(3), against buying and holding and against the slow rule alone.
• Also reported: the median excess return in the 21 sessions after each of the four states, per market and pooled.

TM-H4: do index returns persist? (variance ratios)

• Measure: Lo and MacKinlay’s variance ratio VR(q) on weekly log excess returns, q = 2, 4, 13 and 26 weeks, with the heteroskedasticity-robust z, per market, post-launch.
• Finding of persistence: VR(4) or VR(13) above 1 and significant after Benjamini-Hochberg at 5% on at least a quarter of markets. Finding of mean reversion: the same with VR below 1. Otherwise: no reliable persistence at index level. This decides how the evidence page frames every trend rule; it adds no signal.

TM-H5: a Markov-switching regime beats the 30/70 band

• Model: two-state Gaussian Markov switching on weekly excess log returns (Hamilton filter, EM), refit each January. Hold the market while the filtered probability of the higher-mean state is above 0.5 at the week’s close; cash otherwise; trade at the next session’s close.
• Comparison: long or cash on the existing 30/70 regime (hold in an uptrend).
• Pass: median across post-launch markets of Sharpe(Markov) − Sharpe(band) above zero with its bootstrap interval above zero; both halves; neighbourhood (threshold 0.4, 0.5, 0.6) all positive; deflated Sharpe above 0.95. Expected to fail; if it fails, the band stays.

TM-H6: what followed past trends (tests the trend tab’s descriptive table)

• Pooled across the investable markets, post-launch: the mean excess return over cash in the 13, 26 and 52 weeks after an uptrend began minus the same after a downtrend began.
• Inference: bootstrap over calendar months (each resampled month brings every trend that began in it).
• Pass: difference above zero with 95% intervals above zero at 26 and 52 weeks, in both halves. If it fails, the trend tab’s caveat stands and the finding goes on /trend/evidence/.

What a passing model gets

A methods note, tests, a place on the trend tab or the evidence page, its limits in one honest line, and the live log from the freeze. A model that fails does not become a signal; a clean negative result goes on /trend/evidence/.

Not tested, stated in advance

No other signals, horizons, weightings or universes as hypotheses. Asymmetric GARCH (leverage) is reported as a sensitivity of RK-H1 only. Anything tried after the first results is exploratory and labelled so.

Addendum, 2026-10-13: TM-H7, what followed simple price events

Added at the owner’s request after the TM-H1 to TM-H6 results above were seen (in particular TM-H6, which asks a related question of the regime). Written and committed before any TM-H7 number was computed.

Events, on each market’s daily level (total return for NSE indices and the G-sec index, price otherwise):

• E1, below the 200-day average: the close is below its 200-session average, after at least 20 sessions in a row at or above it.
• E2, back above the 200-day average: the reverse.
• E3, down 10% from the high: the first close at least 10% below the running high; it re-arms only after a new high.
• E4, down 20% from the high: the same at 20%.
• E5, a new 52-week high: the close is the highest of the last 252 sessions, after 63 sessions without one.

Outcomes from the event’s close: the return over the next 21, 63, 126 and 252 sessions (excess over cash for the test; total or price return on the page), the share positive, and the worst fall in the next 252 sessions (lowest close relative to the event’s close). Base rate: the same outcomes from every session.

Per market (description, on the trend tab): each event type’s count, median outcomes and share positive beside the base rate, the list of events with their outcomes, a small-sample warning below 10 events, and back-calculated history kept apart.

The pooled test. Investable markets, post-launch events and post-launch base days only. For each event type, the mean 252-session excess return after the event minus the mean from every base day (return), and the mean worst fall in the next 252 sessions after the event minus the base mean (risk). Inference by bootstrap over calendar months (each resampled month brings its events and its base days), 2,000 resamples, seed 20261013.

• Pass, for each event type and outcome: the 95% interval excludes zero, the same sign in both halves (split at the calendar middle), and the result survives Holm’s correction across the ten tests (five events × two outcomes) at 5%, with two-sided bootstrap p-values.
• Expected, stated in advance: returns after these events will not differ reliably from the base rate (TM-H6); the risk outcome after E1, E3 and E4 may be worse, since volatility clusters.
• A passing risk result is a description of risk, not a timing signal.

Results log

Implementation notes, 2026-10-13, written before any result was seen

Code: pipeline/tipsheet/compute/trend_v2.py (numpy only); tests pipeline/tests/test_trend_v2.py. Run with uv run python -m tipsheet.compute.trend_v2 <rk1|tm4|tm6|…>; output .cache/derived/trend_v2_<test>.json.

• Returns come from load_markets: levels on NSE sessions, carried forward at most five sessions, so an overseas holiday shows as a zero return followed by the next move.
• EWMA starts from the mean squared return of the first 63 sessions.
• GARCH(1,1) is fitted by a grid over alpha 0.01 to 0.25 (step 0.01) and persistence alpha + beta 0.80 to 0.999 (step 0.01, then 0.002 above 0.98), maximising the Gaussian quasi-likelihood with the long-run variance fixed at the fit window’s mean squared return. Recovers simulated parameters in the tests.
• HAR is fitted only on rows whose 21-session target ends on or before the fit date, so no target overlaps the forecast year.
• SPA: two bugs in my first draft were caught by the tests before any real run (the recentring sign, and a missing √n in the variance-ratio statistic); both fixed.
• TM-H4 needs at least 104 post-launch weeks per market.
• TM-H6 uses the published regime runs (trend_runs.parquet), counts a trend only if it began on or after the market’s launch, and splits “both halves” at the calendar middle of the pooled trend start dates (one pooled sample, so the per-market split does not apply).
• TM-H1, TM-H3 (written before their first run): signals may use pre-launch history as warm-up (a 12-month signal on the first post-launch day looks back into the back-calculation, as the path-quality diagnostics already do); only post-launch returns are scored. A market counts if it has at least 756 post-launch sessions. The variance of trial Sharpe ratios in the deflated Sharpe ratio is taken over the 18 TM-H1 variants (each one’s equal-weight daily excess over buying and holding). Deviation: after-tax results are computed only for a strategy that passes before tax. The volatility-scaled rules change exposure a little every day, like the ensemble, and an after-tax replay of that is neither meaningful for a real investor nor needed to reject a rule that already fails before tax.
• TM-H2, TM-H5 (written before their first run): TM-H2 forms thirds only in weeks with at least six counted members (launch-dated, unknown launches left out); with k = members // 3 in each leg. TM-H5 needs 260 weeks before its first out-of-sample year; EM starts each year from the previous year’s fit; the band rule is scored on the same dates as the Markov rule; its deflated Sharpe ratio uses the TM-H1 trial variance.
• RK-H2 (written before its first run): blocks start every 21st out-of-sample session of each market; RK-H2b uses the past standardised 21-session returns whose window ended on or before the block’s start, and needs 250 of them. The chosen RK-H1 model is the lowest pooled loss among passers, or the simplest one whose interval overlaps it.

Results, 2026-10-13 (data to 9 October 2026)

Raw output: .cache/derived/trend_v2_<test>.json, frozen in docs/research/results/trend_models_v2.json.

• RK-H1, volatility forecasts: fail (near miss). 126 markets, out of sample from 2001. Pooled mean QLIKE: trailing 63 sessions 0.349 (trailing 252: 0.357); EWMA 0.94 0.304; GARCH(1,1) 0.267; HAR 0.323. Median gain over the benchmark: EWMA 0.041 (95% interval 0.002 to 0.110), better on 94% of markets; GARCH 0.085 (0.014 to 0.200), 97%; HAR 0.020 (−0.057 to 0.146), 69%. SPA over the seven variants p = 0.013. Post-launch only: EWMA and GARCH pass. Both halves fails for every candidate: GARCH’s first-half interval reaches −0.017; EWMA 0.94’s first and second halves reach −0.010 and −0.0006. The neighbourhood variants EWMA 0.96 and 0.97 clear every condition, but they were registered as checks, not candidates, so they are not promoted. Volatility forecasts beat trailing volatility on nearly every market, as the literature says, but not with the stability this spec demanded; no forecast becomes a signal.
• RK-H2: not run. It needs a passing RK-H1 model.
• TM-H2, 52-week high across sectors: fail. Top third minus bottom third 0.124% a week (Newey-West t 1.46; interval −0.037 to 0.296), 0.223 in the first half and 0.025 in the second. Broad and strategy indices: 0.027% (t 0.57).
• TM-H3, slow and fast blend: fail. 79 markets with three post-launch years. Median Sharpe gap against buying and holding −0.07 (−0.33 to 0.21), ahead on 34% of markets; against the slow rule alone +0.10 (−0.05 to 0.24), ahead on 77%. Median excess return in the next 21 sessions by state: bull 0.85%, rebound 0.93%, bear 0.76%, correction 0.38%. Bear states did not precede negative returns here.
• TM-H4, variance ratios: no reliable persistence. 107 markets, post-launch weekly excess returns. Median VR(4) 1.06, VR(13) 1.13; significant persistence after Benjamini-Hochberg on 3.7% of markets at 4 weeks and 0.9% at 13 (25% needed); no significant mean reversion. By family the medians lean above 1 for broad (VR(13) 1.32) and strategy indices, below 1 for the other assets (0.88).
• TM-H6, what followed past trends: fail. 2,406 post-launch trends (1,179 up, 1,227 down) in 116 markets. Mean excess return after an uptrend began minus after a downtrend began: 13 weeks +1.9 points (−0.7 to 4.4); 26 weeks −0.05 (−3.3 to 3.7); 52 weeks +4.0 (−2.7 to 10.8). The trend tab’s caveat stands.
• TM-H1, volatility-scaled trend: fail; the scaling hurt. 79 markets with three post-launch years. Median Sharpe gap against buying and holding: VS-ensemble −0.08 (−0.37 to 0.20), VS-12m −0.19 (−0.49 to 0.06), VS-hold −0.07 (−0.17 to 0.04); deflated Sharpe ratios 0.03, 0.02 and 0.01 (N = 107). Against its own unscaled version, scaling lowered the Sharpe ratio: ensemble −0.066 (−0.12 to −0.014, behind on 91% of markets), 12-month rule −0.05 (−0.11 to 0.00), holding −0.07. All six neighbourhood variants are negative for all three. For context, the unscaled ensemble against buying and holding: −0.02 (−0.29 to 0.25). A plausible reading, not tested: long-only and capped at full exposure, scaling can only cut risk in volatile spells and never add it in calm ones, and in these markets the volatile spells hold many of the sharpest rebounds. This agrees with Cederburg et al. (2020), not with Moreira and Muir (2017).
• TM-H5, Markov switching against the 30/70 band: fail. 77 markets scored (108 fitted). Median Sharpe gap against the band rule +0.007 (−0.23 to 0.25), ahead on 51% of markets; thresholds 0.4 and 0.6 give +0.03 and −0.02; deflated Sharpe 0.01. A first run crashed on a market too short for any out-of-sample year (no probabilities to score); it was skipped and the test rerun unchanged otherwise. The band stays.

Summary. None of the eight hypotheses passed (RK-H2 was not run). Volatility forecasts came closest: better than trailing volatility on almost every market, but not stably enough in both halves of the history. Trend and momentum timing rules on Indian indices, volatility-scaled or not, and the regime they describe, have not beaten buying and holding or the simple band by more than luck. The results are on /trend/evidence/#review.

• TM-H7, what followed simple price events (addendum): fail; nothing survives Holm. 106 investable markets, 214,351 post-launch base days. Mean 252-session excess return after the event minus the base (9.9%): below the 200-day average −3.1 points (−7.4 to 1.4; 876 events), back above +1.5 (−4.2 to 7.2; 649), down 10% −2.4 (−8.1 to 3.4; 490), down 20% +1.9 (−10.3 to 15.3; 206), new 52-week high +5.8 (0.8 to 10.8; p = 0.028; 549 events), positive in both halves. Mean worst fall in the next 252 sessions minus the base (−10.7%): below the 200-day average −2.2 points (−4.6 to −0.2; p = 0.026), negative in both halves; down 10% −2.0 (−5.1 to 0.5); the others near zero. The two nominal results point the way the literature expects (52-week-high momentum; volatility clustering after a break below trend) but fail Holm’s correction across the ten tests (the smallest p needed 0.005). On the trend tab the per-market tables are shown as descriptions with this result beside them.