What happens if you follow the rule?

# The rule, sealed 2026-09-20

model: momentum_9_1 — 9-1 momentum · top 20 · equal weight (version f518dde5d20d)

## What the rule does
- score = total return over the last 12 months, skipping the most recent month (Jegadeesh–Titman 12-1; the skip avoids the one-month reversal)
- eligible = a name whose close is above its own 200-day average and which has 13 months of history
- hold the top 20 eligible names by score, equal weight, capped at 10% each
- regime filter: when SPY closes below its 200-day average the model holds cash
- data guard: a name with a single-day move over 40% inside the lookback is excluded and named — a jump that size is a corporate action or a bad bar, not momentum, until verified
- rebalance every 21 trading days; 10 bps per side assumed on every change
- parameters are the literature defaults, fixed before any of this data was looked at — nothing here was tuned on the backtest
- variant: nine-month lookback (189 bars) instead of twelve — the rolling out-of-sample tuner chose it from 2022 on once it was in the grid (study §1); a canonical horizon, not a fitted value, but the grid was widened after the sweep showed it ahead

## How it is run on paper
- universe: app/universe/seed_2026-07-25.csv — today's index membership, refreshed yearly; a name that leaves is dropped, a name that joins is added on the next refresh
- cadence: a decision every 21 US trading days on the close; nothing is traded between decisions
- fills: at the NEXT session's close, never at the price the decision was made on
- notional: 100,000 USD on paper; no real money follows the page
- fee: 1.50 USD per order, charged on every fill
- the resting bracket: a stop-loss 20% below each entry and a target 40% above it, both placed against the FILL and never moved; the first day the price reaches either one, the whole position is sold
- from when: the bracket applies to every bar from 2026-09-13 onward. A position opened before that day was bought under the rule as it stood then — a stop, no target. The target is a fixed multiple of the fill, so it was already implied at the buy; this seal corrects it onto that position too, armed for every bar from 2026-09-21 onward and never for a bar before it. Its stop leg has been live since 2026-09-13, unchanged. A rule change is sealed before it applies; it is never applied backwards to bars or positions that predate it
- what a daily price can and cannot settle: a day that reached BOTH levels is recorded as the stop, never the target, because a daily bar cannot say which came first; a price that opened past a level is recorded at that open, not at the level, because it never traded there; every sale says which level fired and on what day
- the regime filter, when it says cash, sells everything on the decision day and waits

## What the evidence cannot claim
- survivorship: the universe is TODAY's index membership; names that left the indices during the test window are absent. Both sides share that survivor universe, but the comparison is NOT fully fair: momentum buys the high-flyers that later disappeared, so its relative edge is optimistic too, not only every absolute figure. Point-in-time membership (M1-1) is the fix and is not built
- returns are close-to-close on dividend-adjusted prices (dividends reinvested), before tax; a name whose price stops updating is held at its last price until the next rebalance drops it
- costs: 10 bps per side on every weight change; no slippage model beyond that
- this configuration is a variant of momentum_12_1 chosen on THIS data (its own rules line says which knob and why) — an in-sample selection, not the literature's default; its edge is measured on the same years that picked it

## Changes
- a change to the rule is sealed here before it applies; the version above changes with it