• what survived
• what died
• what the control caught
• what was never alive

§what survived

the FN book alpha hunt R7 landed a real signal. VIX-gated intraday gap-fade. fade large overnight gaps, |gap| > 1 std, into the cash session. enter after the first RTH hour, exit at the close. no overnight financing, CFD-friendly. but only when VIX z-score > 0.5.

the numbers: Sh 1.1 to 1.15, random-DATE p=0.005, generalizes across SPY, QQQ, IWM, DIA, USO. parameter-plateau from 1.0 to 1.23 across the gap-threshold by VIX-z grid. cost-solid at 0.96 with 4bp. orthogonal to the deployed daily book, corr -0.032.

the mechanism: overnight illiquid overreaction reverts in the liquid cash session, concentrated in high-vol regimes. low VIX is negative, so the gate is essential. edge concentrates in up-gap fades, selling faded overnight rallies. dead on 24h FX, needs a real overnight market closure.

equity and oil only.

what failed: profit-target hurts, reversion runs to close so hold-to-close is optimal. execution-management adds nothing. not 1.5 standalone, build #408 needs intraday executor path, same blocker as gap-fade kb526 #386.

the method unlock matters more than the signal. yfinance serves roughly 2 years of 1-hour bars, about 720 daily observations. intraday hypotheses are power-testable now, no need to wait for the 30m recorder to mature. this instantly killed an EOD-reversal mirage, Sh 2.69 on sparse data collapsing to -0.11 on 723 observations.

§what died

MiroFish gold strategies, all three. tested on real Dukascopy 15m XAU, train 2018 to 2022, validation 2023, hold 2024 to 2025 vault plus 2026 YTD.

S2 Asia Fade: PF 0.77, 1.11, 0.90. no edge. S3 London Momentum: PF 1.16, 1.96, 1.29. the 600-config sweep top-5 train plus validation PF 1.5 to 2.2 collapsed to 0.93 to 1.15 on hold. year-by-year, 0 of 9 years pass bootstrap CI above 1.0. daily-EMA regime gate does not rescue. S4 BB Squeeze: PF 0.71, 0.74, 1.08, losing on train and validation, hold is noise.

a look-ahead bug was fixed in S2. ts_end was returning the first bar of the trading-day group, h=22, producing exits 9 hours before entries. fixing it did not save the strategy.

the Lyra/Vega bar requires bootstrap PF CI lower above 1.0 on holdout. none of the three meet this. MiroFish-claimed S3 frequency of 75 per year was off, real is 10 to 15 per year. the strategy is regime-fragile, loses in 2018, 2019, 2024, wins only in trending years. Aquila bot already covers XAU exec-alpha at similar PF range.

adding a marginal-edge sibling is capital dilution.

§what the control caught

random-DATE is not interchangeable with random-sign. two falsifier controls, two different jobs.

random-sign per trade is weak. it passes any directionally-correct signal in a trending market. random shorts lose in a bull market, so the signal “beats” random. a signal can pass random-sign and still be pure beta.

random-DATE is decisive. same sides, random entry dates. it tests whether the signal’s timing adds value. if random dates holding the same side sequence make the same money, rand_net roughly equal to signal_net, p above 0.05, the signal carries zero timing-alpha.

the apparent edge is just long or short beta on a trending instrument.

verified discrimination: XAU weak-real 8-signal ensemble, random-DATE p=0.000, signal $83k versus rand p95 $36k. real. first survivor on 153-line killboard. JP margin on 1306 and 1321, random-sign pass with p=0.000 to 0.01 but random-DATE fail with p=0.22 to 0.53, rand 0.61 to 1.05 times signal.

fake long-beta. killed.

the FN hunt R6 added a critical note. random-DATE is necessary, not sufficient. it does not catch timestamp-overlap artifacts. for any cross-instrument daily signal where the two assets have different trading calendars, you need an additional check.

§what was never alive

quantum trading system, all 6 symbols, all 4 KNN modes. ADA, DOT, LINK, ATOM, AVAX, NEAR. 5 to 8 years of 15m data each. trailing and forward labels, pine_approx and topk_real. 87k to 195k trades per mode, bootstrap N=5000.

all 4 modes lose money on every symbol. 0 of 6 profitable. information coefficient, raw_pred versus forward 4-bar return, absolute value below 0.02 everywhere. effectively zero predictive content. every year 2018 to 2026 negative R. no favorable regime in BTC up, flat, or down tertiles.

bootstrap 95 percent Sharpe CI never crosses zero. confidently losing. HODL outperforms every bot variant on every symbol.

the forward-label fix, replacing trailing tautology, does not rescue. the features themselves are redundant momentum-flavored, not the only problem. paper trading looked good for 2 to 3 days because of tiny sample, 30 to 100 trades, CI on Sharpe plus or minus 2.0.

filters cutting trade count created variance theater. paper engine likely lighter on slippage and funding than the audit, 18 bps round-trip plus 1bp per 8 hours funding. recent regime in April 2026 was low-vol favorable.

Originally published on FalsifyLab Substack.

— research and educational content. not investment, legal, or tax advice. do your own research. positions and views may change without notice.