What this note is
A worked example of the bar set out in Walk-forward is the report (18 October 2026), applied to a deliberately simple cross-sectional momentum rule.
Results in this draft are placeholders. Francis has not authorized a production backtest run for this article. No Sharpe, return, drawdown, or capacity number below is real. Prefer a blank table to a fake chart.
This is a community research note. It is not investment advice. It is not a live allocation, a product, or a recommendation.
Hypothesis (one paragraph)
In a liquid US large-cap equity universe, a plain 12–1 month cross-sectional momentum sort — long the top decile, short the bottom decile, monthly rebalance — retains a positive after-cost edge when evaluated under purged walk-forward folds that never tune on the test window, under 2× slippage stress, and under a one-bar clock shift. If the edge lives in only one subperiod, one sector, or one parameter cell, the note ships as a negative or as a labeled single-regime curiosity — not as a book.
Rules (toy spec)
| Piece | Spec for this case |
|---|---|
| Universe | US common stocks, point-in-time membership in a large-cap index (exact index TBD in the authorized run) |
| Signal | 12-month total return, skip most recent month (12–1) |
| Portfolio | Equal-weight long top decile / short bottom decile |
| Rebalance | Month-end close → next open (or next close; freeze one convention before the run) |
| Costs | Round-trip haircut as % of ADV; baseline + 2× stress (exact bps TBD from the desk cost model) |
| Neutralization | None in v0 (sector-neutral is a listed sensitivity, not the base case) |
| Gross exposure | Written in dollars and in shares at the volume the desk can actually hit — number TBD; not invented here |
Codex may draft the feature function against the existing harness. It does not pick the window, drop a name, or rerun until the chart behaves.
Walk-forward / purge design
| Check | How this case implements it |
|---|---|
| Walk-forward | Rolling train / test folds. Test window is never used for lookback length, decile count, or cost assumptions. |
| Purge / embargo | Embargo at least one rebalance gap between train and test so overlapping 12-month returns do not leak labels into the test fold. Exact embargo length frozen before the run. |
| Parameter plateau | Neighboring cells (e.g. 9–1 / 12–1 / 15–1; deciles 8/10/12) must be scored. One green cell is not a pass. |
| Clock shift | Rebalance bar moved by ±1 trading day; sign of after-cost mean return must not flip. |
| Capacity | Max participation vs ADV written down; if the edge needs more size than the desk can hit, label it research-only. |
A rule that flips because Tuesday became Wednesday was never a rule. Same for a Sharpe that lives in one ticker.
Cost stress
- Baseline slippage / commission from the desk cost model (cite the model version in the authorized run output).
- 2× haircut: after-cost mean and hit rate must remain non-negative (or meet the pre-registered hurdle) under double costs.
- If the edge dies at 2×, the note ships as a negative — silent discard is how a desk talks itself into a curve.
Falsification criteria (pre-registered)
Write these before looking at OOS numbers:
- After-cost mean monthly return ≤ 0 in the pooled OOS folds under baseline costs → fail.
- Sign flip under 2× cost stress → fail.
- Sign flip under ±1-bar clock shift → fail.
- Only one neighboring parameter cell green → fail (or publish as fragile, not as a book).
- Edge concentrated in a single calendar regime (e.g. only 2017–2019) → first-paragraph disclosure; do not market as a general momentum book.
- Edge concentrated in one name or one sector after a simple leave-one-out → label single-name / single-sector; stop calling it a book.
If a note cannot say what would make it wrong, it does not get published.
Results — framework only (numbers TBD)
| Metric | Authorized run | This draft |
|---|---|---|
| OOS mean monthly return (after cost) | pending Francis run | TBD — not invented |
| OOS Sharpe (after cost, stated RF) | pending | TBD |
| Max drawdown (OOS) | pending | TBD |
| 2× cost stress: sign / mean | pending | TBD |
| Clock-shift sign | pending | TBD |
| Parameter plateau (cells green) | pending | TBD |
| Capacity at stated participation | pending | TBD |
Pass / fail against the falsification list: pending authorized output.
When the run lands, replace this section with the frozen report table and a human-edited paragraph that points at the table. Model prose that cannot point at a table does not ship. If the run fails a row above, the note still ships — as a negative.
How it is filed
Hypothesis, universe, cost model, OOS window, and invalidation live with the other research notes. We do not publish a live allocation from this page. We do not publish position sizes.
Not investment advice. Community research viewpoint only. Framework draft — no backtest numbers until an authorized run is attached.