Data to 20 November 2026
About

Pre-registered research

An Indian model library: what I can learn from Allocate Smartly, and a proposal

Written 2026-10-03 for the model audit (docs/briefs/model_audit.md). This is a proposal for the owner’s decision, not a pre-registration. Each study in it still needs its own spec before anything is computed.

What Allocate Smartly does (read from its public pages, 2026-10-03)

Allocate Smartly (allocatesmartly.com) replicates published tactical asset-allocation strategies for US investors. Its strategy pages, screener and “current allocation” views sit behind a members’ login. Everything below comes from the public pages cited, and where the public pages are silent I say so. Nothing here is copied text or design.

What tipsheet already does that this site does not (on its public pages)

Allocate Smartly’s discipline is mostly in the plumbing: one asset set, one cost model, timing robustness by default. Mine is mostly in the inference. The proposal combines the two.

Proposal: the tipsheet model library (India)

  1. One engine, one footing.
    • Engine: every tactical model (trend rules, the trend ensemble, sector and index momentum, the lab’s rule-based portfolios) runs on lab/engine.py: lots, costs by date, tax by financial year, and trades at the next close after the signal.
    • Fixed instrument set: Nifty 50, Next 50, Midcap 150 and Smallcap 250 index funds; gold ETF; 5-year G-sec (until a long G-sec TRI arrives); liquid fund. Pre-launch history is the tracked index minus a real expense ratio, flagged on the chart.
  2. Every model carries three dates:
    • its source date (when the rule was published, for example Faber 2007);
    • its entry date (when it entered my library: its spec commit);
    • its data start.
    • Charts shade three periods differently: before the source date (in-sample for the author), between the two dates (out of sample for the author, but seen by me), and after the entry date (truly out of sample for me).
    • A table gives the return after each date beside the full-history return. This is the McLean and Pontiff (2016) test, model by model.
  3. “What each model holds today” board.
    • One row per model: today’s target, the date of the last change, and how close the signal is to switching.
    • A “changes likely at this month-end” flag on the last three sessions of the month.
    • An aggregate row: the average equity exposure across all models.
    • Descriptive only, beside each model’s verdict (a deflated Sharpe that passed or failed).
  4. Timing luck by default. Every monthly model is shown with its spread across 21 trading-day offsets (the trend barometer already measures this) and as a tranched version. The headline number is the tranched one.
  5. Blends.
    • Readers can combine library models with fixed weights; the blend is computed on the same engine, with tax.
    • No optimiser, and no parameter changes: the model list is the menu.
    • A blend starts only when all its models have data.
  6. A “didn’t survive” shelf. Retired or failed models stay listed with the reason and the evidence (for example sector momentum v1: no edge over equal weight, deflated Sharpe 0.41).
  7. Live tracking. From each entry date, the pipeline appends the model’s daily return to a frozen log that never recomputes. This gives a real out-of-sample record, separate from the backtest, which does get recomputed when data are repaired.

First models to load (all pre-registered or published in the literature)

Each needs its own spec, with the Indian substitutions written down before it is run.

What it needs from other lanes

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.

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