dennisthiessenandClaude Opus 5.5
231372fb3a
- The breakeven study grades at the shadow book's capacity AND risk per trade, and labels each with where it came from. A missing setting row now reads "shadow_book default" instead of crediting the setting; the committed report said shadow_book_capacity though the snapshot has no such row. - _production_book_config reads strategies through the same _portfolio_monitor_strategies() hook as the monitor. - breakeven_gate_reset is rejected when there is no post-stop episode or cooldown for it to start, and the study skips its diagnostic arm for a book without one, so a production config change cannot abort the run. - identical_to_control is renamed ties_control_on_graded_metrics and needs real values: be_2r ties on every graded metric but still changed 2 trades. - _BookConfig drops an unused field and hashes by identity; the min-R:R sweep swaps its gate in with dataclasses.replace instead of passing it. - New tests cover the parity guard's success path, the no-config fallback and the gate-reset check. The breakeven report is regenerated: every row, diagnostic row and the parity check are identical; only the labels, the risk fields and the renamed flag changed. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
4.7 KiB
Residual normalization — findings
2026-09-08: closed negative at the signal screen. Keep production unchanged. The proposed normalization did not meet the rule frozen in the experiment protocol, and its point estimates did not improve on the current momentum signal. No portfolio replay was run: the protocol requires a passing signal screen before that stage.
The candidate used the production SPY-adjusted 12-1 return sum divided by sample standard deviation of the same 231 daily adjusted returns. All comparisons used identical as-of dates and 30-SPY-session forward-return endpoints. No new factor, lookback, volatility weight, sizing rule, or exit was tested.
| Era | Non-overlapping windows | Current signal, matched coverage | Normalized signal | Difference |
|---|---|---|---|---|
| Full period | 73 | 0.01446 | 0.01284 | -0.00162 |
| Before 2021 | 30 | 0.00776 | 0.00496 | -0.00280 |
| From 2021 | 42 | 0.02377 | 0.02259 | -0.00118 |
Values are mean Spearman information coefficients (IC): how well signal ordering agrees with subsequent stock returns. They are neither Sharpe ratios nor trading returns. Higher is better for this positive-momentum hypothesis. The candidate needed full-period IC >= 0.03, at least 12 windows, and positive mean IC in both eras. It passed the count and era-sign conditions but failed the 0.03 threshold. The rule was evaluated before rounding.
The paired normalized-minus-current IC difference has a 95% moving-block interval of [-0.00496, +0.00156]. This spans zero: the study does not establish that normalization is worse, and it supplies no convincing improvement evidence. The mean top-minus-bottom-quintile 30-session return spread also fell, from 0.967% to 0.751%. That is an unconditional diagnostic, not the P&L of the long-only, gated, capital-constrained production book.
Coverage and checks
- The snapshot completion manifest matches 4,650 registry tickers, 5,081,073 OHLCV rows, and 4,144 research-only names. The experiment used the remaining 506 production-universe registry names, averaging 485.4 paired stocks per retained window.
- Actual retained as-of dates run from 2017-01-06 to 2026-05-01; the final forward endpoint is 2026-06-15. Later dates cannot necessarily add another complete non-overlapping 30-session window. SPY history starts in January 2016, so a full one-year signal cannot start in July 2016.
- One retained window crosses the 2021 boundary and belongs to the full-period result only, explaining why the era counts sum to 72 rather than 73.
- The unrestricted production control had 35,441 retained stock/date cells; paired control and treatment each had 35,433. Its mean IC was 0.01447, essentially the same as paired control. Coverage restriction did not account for the observed result.
- Numerator parity was checked against the production implementation on each eligible evaluated cell. The snapshot and frozen source hashes were unchanged across the run. SQLite was opened read-only.
- Twelve targeted tests passed, including numerical invariance, production parity, lookahead exclusion, zero-variance handling, missing-data pairing, calendar overlap, and an end-to-end screen check. Ruff passed for the runner and its tests. The existing Python runtime was usable outside the filesystem sandbox; no environment replacement was needed.
- A 2026-10-10 code review added four tests (missing prices, file release, paired-coverage guard, source ids) with the runner fixes they cover. The report was regenerated with the fixed runner; its screen results are byte-identical to the original run.
Interpretation and decision
My proposed mechanism was plausible but unsupported by this experiment. Do not add normalization to production or search denominator variants on this history. This ends this specific experiment, not all possible residual-momentum research.
The current signal also falls below 0.03 on this particular long-history screen. That cautions against treating its predictive evidence as universally strong; it does not by itself measure or invalidate the complete production book. The unconditional universe, true session spacing, dates, and return objective differ from earlier setup/portfolio studies. Historical constituents are still projected backward, this history has already been inspected, and the bootstrap intervals remain exploratory. No new CAGR, drawdown, cost-stress, or live result is claimed.
Artifacts: machine-readable results, generated tables, runner, and tests.