Skip to content

WORKED_EXAMPLE.md

Latest commit

 

History

History
43 lines (33 loc) · 3.43 KB

WORKED_EXAMPLE.md

File metadata and controls

43 lines (33 loc) · 3.43 KB
Raw

Worked Example: Process-Conditional Verdict

This is a constructed, isolated illustration of Penrose's deflation gate running inside the real P7/P8 pipeline. It is not alpha, not a real strategy, and not evidence of a profitable trading rule. The same byte-identical return series R is scored twice through run_backtest(...) and stages.p8_verdict(...); only the disclosed trial ledger changes.

By construction, R is tuned to sit just inside the boundary: Process B's deflated DSR lands just below the kill threshold while A stays above it, so the flip is deliberately a near-threshold case (a small change to the effect size would move B back across the line). This is intentional - the point is to isolate deflation as the only gate that differs, not to claim a wide margin.

Full verdicts include more than DSR: minimum OOS bars, 3-fold sign stability, regime fragility, bootstrap edge CI, permutation alignment, walk-forward consistency, holdout confirmation, and provenance caps. In this construction those non-deflation gates pass for both processes; Process B is killed only after the ledger deflates the OOS-slice DSR.

Same byte-identical return series:

  • SHA-256: a4c125c10ab55b515612347f0185523df6fe065b6e0d9d950ca675ff055b19cd
  • Length: 4000
  • OOS bars: 1200
  • Full-series annualized Sharpe, descriptive only: 0.873098
  • DSR thresholds: kill below 0.9, watch band (0.9, 0.95)
Process Search lineage Series hash Bars/year n_oos n_trials ledger sr_variance OOS PSR OOS DSR Real P8 verdict kill_reason
A: preregistered one hypothesis 1 a4c125c10ab55b515612347f0185523df6fe065b6e0d9d950ca675ff055b19cd 252 1200 10 0.000000 0.999700 0.999700 watch None
B: selected best of 200 200 a4c125c10ab55b515612347f0185523df6fe065b6e0d9d950ca675ff055b19cd 252 1200 209 0.000500 0.999700 0.897600 kill no_oos_edge

Gate output:

  • A folds: {'folds': [0.128, 2.199, 3.082], 'consistent': True}
  • A regime: fragile=False n_partitions=9
  • A bootstrap: edge_ci=[0.00054, 0.00148] includes_zero=False
  • A permutation: p_value=0.0005
  • A walk_forward: consistent=True (per-window Sharpe magnitudes are an artifact of the synthetic walk-forward frame, not R's full-series Sharpe; the gate only checks consistency)
  • A holdout: {'holdout_sharpe': 0.798, 'holdout_psr': 0.9221, 'nbars': 800}
  • A p8 rationale: costs/capacity are MODELED placeholders - capped at watch until measured (E2)
  • B folds: {'folds': [0.128, 2.199, 3.082], 'consistent': True}
  • B regime: fragile=False n_partitions=9
  • B bootstrap: edge_ci=[0.00054, 0.00148] includes_zero=False
  • B permutation: p_value=0.0005
  • B walk_forward: consistent=True (per-window Sharpe magnitudes are an artifact of the synthetic walk-forward frame, not R's full-series Sharpe; the gate only checks consistency)
  • B holdout: {'holdout_sharpe': 0.798, 'holdout_psr': 0.9221, 'nbars': 800}
  • B p8 rationale: OOS score 0.898 (<0.9), edge_t 3.43

Why the verdict changes:

  1. A normal backtester sees only R; Penrose also sees the search lineage that produced R.
  2. P7 computes PSR/DSR on the OOS slice, not the full series, and adds regime partitions to the ledger trial count.
  3. The returns, dates, bars/year, costs, robustness inputs, and holdout are unchanged. Process B's 200-trial ledger raises the deflation denominator and cross-trial Sharpe variance, dropping only B below the P8 DSR kill threshold.