Independent validation of the overnight sleeve book:sleeves are real,are redundant
Date: 2026-07-25 · Type: Independent reproduction of shipped trading claims (#2940), on real price history. Read-only; no orders. Artifact: overnight_sleeves_validation.py → results Data: Yahoo daily OHLC, full history — SPY 8,428 bars from(33.5y), QQQ 6,886 from 1999, IWM 6,578 from 2000, GLD 5,453 from 2004.
Why this was needed
PR #2940 ships an overnight book with specific quantitative claims — capitulation "+18bp/night, t=3.5 over 30y", a bear-rally fade "−13bp (t=−2.1)", a book at "Sharpe 1.78, max DD −3.1%". No test in that PR asserts any of those numbers, and neither the cited backtests (oracle ledger downtrend-decomposition-*, bandits-trader-*) nor any experiment reproducing them exists anywhere in the repository. The claims were unreproducible from the repo. This reproduces them from scratch, with the gate math ported verbatim from lib/overnight-trader.js (parameter lock: no tuning, no added filters).
The baseline nearly everyone forgets
Unconditional SPY overnight return: +3.27bp, t = 4.49, n = 8,427.
This is the well-documented overnight anomaly — the equity premium accrues disproportionately outside trading hours (Lou, Polk & Skouras, A Tug of War: Overnight versus Intraday Expected Returns, JFE 2019, CRSP+TAQ). Our 33-year measurement independently confirms it exists.
Every sleeve must beat +3.27bp, not zero. A sleeve that returns +5bp overnight has not found an edge; it has found a complicated way to be long at night.
Result: excess return over simply holding overnight
| sleeve | raw bp | its baseline | excess bp | t | p |
|---|---|---|---|---|---|
| capitulation SPY | +13.59 | +2.47 | +11.12 | +2.40 | 0.017 |
| bear-rally fade SPY | −8.64 | +3.92 | −12.56 | −2.11 | 0.035 |
| uptrend notflat SPY | +5.00 | +2.81 | +2.18 | +1.44 | 0.15 |
| uptrend notflat IWM | +5.40 | +4.41 | +0.99 | +0.42 | 0.67 |
| uptrend notflat GLD | +4.37 | +4.55 | −0.18 | −0.07 | 0.95 |
| uptrend flat QQQ | +4.35 | +5.50 | −1.15 | −0.57 | 0.57 |
The two conditional sleeves are real signal
Both survive the correct baseline. The fade is the striking one: measured t = −2.11 against a claimed t = −2.1 — the PR's own number, reproduced to two significant figures on independent data. Capitulation reproduces at +13.6bp against a claimed +18bp (76% of claim, same sign, same order of magnitude), and it is not decaying: first half +8.19bp, second half +18.98bp.
The four uptrend sleeves add nothing
All four excesses are statistically indistinguishable from zero, and two are negative. GLD is −0.18bp — literally nothing. These sleeves are an elaborate mechanism for capturing the unconditional overnight premium that a plain "hold overnight" captures for free.
They are not harmful — they do earn +4 to +5bp — but the trend and vol-regime gating buys no measurable improvement over holding. With allocPct spread across whichever sleeves fire, four sleeves that add nothing dilute allocation away from the two that do.
Honest limits on the positive result
- Neither real sleeve clears the multiple-testing bar. Harvey–Liu require |t| > 3.0 for a new
factor claim; capitulation excess is t = 2.40 and the fade is t = −2.11. Against a pre-specified hypothesis those are respectable; as two of six tested sleeves they are not decisive.
- The book's headline is still unreproduced. Nothing here validates "Sharpe 1.78, max DD −3.1%".
That requires a full portfolio backtest with position sizing and costs, which does not exist in the repo.
- Costs are not modelled here. These are gross overnight returns. A +11bp excess is real but
thin — commissions, spread at the open auction, and slippage are unmodelled and could consume a large share of it. The published guidance that transaction costs commonly eat 20–40% of a backtested edge applies directly.
- Yahoo daily data, not survivorship-adjusted CRSP. Fine for index ETFs (no survivorship issue)
but the open prints are as-reported, not audited.
What follows
- **The
edgeMinNchange in #2947 is
corroborated.** The fade at t = −2.11 sits below the |t| > 3.0 bar — exactly the reasoning for raising its live-evidence requirement to ~330 nights.
- The uptrend sleeves should be justified or dropped. They need to demonstrate excess over
holding overnight, which onyears they do not. That is a proposal, not a change made here.
- Cost modelling is the next blocking measurement for the two real sleeves — an 11bp gross
edge may or may not survive the open auction.