Model dashboard

A shadow is a second frozen model on the same universe, scored every night in its own run, into its own workbook and its own rows. It trades nothing. It exists so that a candidate the backtest already chose accrues a forward record before anything switches to it — the live year is there to stop a materially worse model and to prove the lane runs, not to prove a better one.

Promotion is by process, never by profit and loss. The first evaluation is taken when the two models have 250 days both committed in advance, and every 125 paired days after it. It promotes only if the shadow's live ranking of the implementable return is below the production model's by no more than 0.002 in rank correlation, if the two models' scores correlate at least 0.30 at that first evaluation and at least three-quarters of the first evaluation's own value at every later one — a drift check, not a second level check — and if the lane has no coverage holes of its own. Each book's own return is reported at those evaluations and decides nothing.

Decision rule fixed in docs/pre-registered-S.md, before the first shadow night. the registrations →

The dashed amber stretch was reconstructed in arrears after the scheduler failure and after a week without network — real accounting over a real book, but the positions were never committed to in advance, so it demonstrates plumbing, not foresight. The solid green stretch carries the live stamp, and only the days inside it that were committed before the outcome existed are evidence. Turnover is charged at 3 bps round-trip; the newest day is unmarked until its forward return is realized, so the account is valued at the last marked day.

Left chart: the full scored cross-section (up to ~500 names a day), every scored day. Each name with a realized next day falls into a rank decile — the model's least favorite tenth on the left, its favorite tenth on the right — and each bar is the realized next-day excess return, averaged within each day first, then across days, with a block-bootstrap SE. If the model orders stocks at all, bars should rise left to right. Note which end separates: a long-only book only harvests the top decile, so separation at the bottom alone — the rejects underperforming — would not be captured by this construction.

Right chart: the same question one day at a time — each point is that day's rank correlation (IC) between model ranking and realized outcomes. The backtest's frozen-holdout IC was about +0.01: on any single day the correlation is almost invisible, which is why it is only usable across ~500 names at once, and why a month of days decides nothing.

One cumulative line over the frozen model's holdout. Implementable is what an execution one day behind the signal earns: enter at the next open, hold through the following night, exit at the open after. It is a sector-excess return centered across the day's cross-section, so the average stock is zero. The verdict is fixed in docs/pre-registered-P3.md; costs at the open are not measured, so the real number is lower by whatever the spread and impact cost at the open.

Two other legs were measured and are not reported here: the credited close-to-close return the record books, and the intraday-only leg (next open to that same close) that P1 decomposed out of it. Neither is an execution that can follow an 18:30 signal, so neither is a result. Both remain in backtest_daily and in the Phase P registration — the audit trail keeps what was measured even when the browser stops publishing it.

What the +4.93 is made of. The tilt: roughly half of it is the return to the book’s beta tilt — slope +0.42 on the forward market, 52% of the mean on this decade, residual +2.36 at t 1.45 (AG-018, an agent-fleet finding, not a registered block); it is never restated beta-adjusted, here or anywhere else on this site. The timing: 87% of the implementable return sits in the close-to-open leg on 2016–2021; since 2022 the two legs split evenly and neither is distinguishable from zero (AG-013). The ranking: decile 1 earns +5.43, decile 2 +1.22, deciles 3–10 sit at zero, and inside decile 1 the five sub-buckets are not ordered (R3, registered).

The live column is the only part that is evidence, and only once there are enough days committed in advance for it to mean anything.

backtest The frozen model scored on dates it never traded, after the fact. A description of the holdout, not evidence — only live days are evidence.

Every figure is read off the implementable leg the timing panel reports: the next session’s open to the open after, each name’s return over its own sector’s median, the book averaged per day. Yearly figures assume 252 trading days: the yearly return is the daily mean × 252, the yearly volatility the daily standard deviation × √252, and the Sharpe ratio is the first divided by the second. No cash rate is subtracted. The return is already measured against the stock’s own sector, so it is a spread rather than money that could have sat in cash; a fund would call this Sharpe ratio an information ratio.

The ± on the Sharpe ratio is the site’s block bootstrap — 2,000 resamples of five-day blocks, the same draws that give the mean its SE in the timing table — applied to the ratio itself, and it is undefined below 6 days. The Sortino ratio divides by the downside deviation instead: the root-mean-square of the losing days, with every day in the count. Drawdowns are measured on the summed series, the cumulative line in the timing panel, from a starting level of zero. The worst-5% average needs at least 20 days.

Costs at the open are not measured, so the switch lets you choose one. It charges the round-trip cost on the share of the book the live record has actually replaced each night, as the same amount every day: a cost lowers the return and the ratios and deepens the drawdowns, and leaves the volatility unchanged. The breakeven is the cost at which the whole return is gone. The charts stay before costs.

The paper account above is marked on the record’s close-to-close return, which no execution following an 18:30 signal could obtain; the site stopped reporting it as a result on 2026-09-12. No Sharpe ratio is computed from it.

backtest The frozen model scored on dates it never traded, after the fact. A description of the holdout, not evidence — only live days are evidence.

Both books are scored on the same evening, on the same cross-section, and both are evaluated on the same leg the rest of this page reports: the next session's open to the open after. The comparison is paired — the difference is taken date by date and its standard error is bootstrapped from the difference series, not from the two books' own standard errors. Combining those would overstate the uncertainty of a difference between books holding overlapping names by several times.

Only dates committed before their outcome existed are paired here. A day scored in arrears is not a commitment, so it cannot settle a comparison between two commitments, whatever its stamp says.

Turnover is the share of yesterday's book gone today, over consecutive scored dates. The breakeven beside it is R2's convention — the whole gross divided by that fraction, the round-trip cost in bps that would eat it, as +4.86 bps on 20.2%/day gave ~24 bps there — and it is gross of a cost nobody has measured: costs at the open are not in any number on this site.

The power line says how many paired days a difference would need before its t reaches 2, at the dispersion AG-020 measured on 2,685 dates. Both answers are decades. That is the point: the companion is worth running because it commits the comparison in advance, not because the comparison will resolve. AG-020 is an agent-fleet falsification, not a registered block.

Each bar is the book's mean implementable excess (bps/day) on holdout days in one market state, with a block-bootstrap SE and the day count. States are fixed in advance: SPY's trailing 21-day volatility in expanding point-in-time terciles, SPY above or below its 200-day average, and the cross-sectional dispersion of members' trailing 21-day returns in terciles.

The table is the registered P2 test — top state minus bottom state against 200 within-date permutations with the book fixed, three tests under Benjamini–Hochberg. It was run on the credited close-to-close leg, not on the leg these bars show. Its p and q are reported because they are the registered result; its observed difference and null percentile are not, because they are in credited bps and would read as belonging to the bars. A regime claim about the implementable leg needs its own registration before it is run — re-running the test on whichever series looks better is exactly the forking path the pre-registration rule exists to prevent.

The permutation null asks whether a random book of the same size would show the difference. It does not carry the book's day-to-day swings, so a verdict of “regime-dependent” is narrower than the bars might suggest.