StockpickerAiStockpickerAi

โš—๏ธ Binary Lab โ€” Reproducibility Study

Eight experiments. Five rungs. One engine demolished, one audit framework earned. Run any experiment live, then read the writeups. Everything runs against a frozen 9-symbol OHLCV snapshot bundled with the study โ€” no API calls, no credentials, so the results are reproducible bit-for-bit.

Data: 9 symbols ยท frozen at 2026-06-18 ยท engine = binary_exhaustion_engine ยท verdict = z โ‰ˆ +0.6 vs matched-bias null (null).

The 5-rung audit ladder

  1. Rung 1 โ€” Wrong payoff. The original strategy "won" in shares-only PnL but lost in realistic option-premium PnL. (EXP1, EXP3)
  2. Rung 2 โ€” Sampling noise. A single coin-flip control arm has its own variance. Sweep 200 seeds to characterize the null. (EXP4)
  3. Rung 3 โ€” Wrong null. A long-biased strategy beats a fair-coin null structurally. Match the null's bias to the strategy's. (EXP5)
  4. Rung 4 โ€” Autocorrelation. 494 fires of overlapping 10-day windows aren't 494 independent observations. Filter to non-overlap. (EXP6, EXP7)
  5. Rung 5 โ€” Untested detector. A framework that has only ever said "no edge" might be too conservative. Feed it a known oracle. (EXP8)

Verdict on the engine

  • z = +0.6 to +1.0 vs matched-bias null at every honest test.
  • Apparent z = +3.83 at T+10 collapsed to z = +0.60 once autocorrelation was removed.
  • Worst 20-bar drawdown windows: engine z = โˆ’1.25.
  • No detectable directional skill in any framing or regime.

Verdict on the framework

  • Detects 55% synthetic oracle at z = +2.50 on n = 675 โœ“
  • Nulls a 50% oracle at z = โˆ’0.16 โœ“ (no false positive)
  • Subtracts long-bias from a no-skill biased oracle โœ“
  • Two-way meter. Bedrock.