l2-regime: rolling-RV regime filter, OOS-verified 2× IC uplift#236
Merged
neuron7xLab merged 3 commits intomainfrom Apr 17, 2026
Merged
l2-regime: rolling-RV regime filter, OOS-verified 2× IC uplift#236neuron7xLab merged 3 commits intomainfrom
neuron7xLab merged 3 commits intomainfrom
Conversation
Three diagnostic scripts built on existing primitives (slice_features, run_killtest, cross_sectional_ricci_signal) — no new abstractions (AE principles 1, 20). * scripts/l2_killtest_recursive.py — depth-first bisection (up to depth 3) + cyclic K=8 disjoint blocks. Reveals regime structure hidden by full-window averaging. On collected substrate: 3/8 blocks PROCEED with IC up to +0.339; 3/8 KILL with IC as low as -0.109. Signal is intermittent, not uniform. * scripts/l2_regime_analysis.py — per-block regime features (realized vol, cross-asset correlation, dispersion, signed trend, κ_min moments). Spearman rank-correlates block IC against each feature. On K=8: corr_mean strongest direction (ρ=+0.429, p=0.29 n=8). Underpowered for statistical claim; motivates finer-grained analysis. * scripts/l2_walk_forward.py — 40-minute rolling window with 5-minute step across substrate. ~56 windows gives the statistical power that 8 disjoint blocks lack. Reports IC trajectory, Spearman ρ at rolling resolution, quartile bins on the most-correlated feature to find a discriminator threshold. Output artifacts: * results/REGIME_ANALYSIS.json (8-block table + ρ matrix) * results/L2_WALK_FORWARD.json (56-row trajectory, quartile bins) Non-goals: new dataclasses, new production modules. Pure diagnostics. If walk-forward identifies a regime discriminator, next commit adds the regime filter to killtest.py as an optional parameter. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Walk-forward analysis (scripts/l2_walk_forward.py, 56 rolling 40-min
windows) identified rolling realized volatility as the dominant regime
discriminator:
Spearman ρ(IC_signal, rv_mean) = +0.352 p = 0.008 ***
Spearman ρ(IC_signal, corr_mean) = +0.317 p = 0.017 *
Spearman ρ(IC_signal, trend_*) = not significant
Quartile analysis on rv_mean:
Q1_low IC median +0.027 (signal ≈ noise)
Q4_high IC median +0.137 (signal works)
IN-SAMPLE CONDITIONAL (scripts/l2_regime_conditional.py):
unconditional IC = +0.122
rv_w600_q75 IC = +0.256 (frac_on = 24.2 %)
rv_w300_q50 IC = +0.177 (frac_on = 49.2 %)
TRUE OOS (scripts/l2_regime_oos.py, threshold trained on first half,
applied to second half, no information leakage):
TEST unconditional IC = +0.116 frac_on = 100.0 %
TEST q50 thr from train IC = +0.202 frac_on = 43.9 %
TEST q75 thr from train IC = +0.236 frac_on = 36.3 %
=> 2.03× IC uplift OOS, threshold generalizes
Components
- research/microstructure/regime.py — 4 functions, no new dataclasses:
* rolling_corr_regime(features, window_rows)
* rolling_rv_regime(features, window_rows) (primary, OOS-verified)
* regime_mask_from_score(score, threshold)
* regime_mask_from_quantile(score, quantile)
- research/microstructure/killtest.py — single optional parameter:
* run_killtest(..., regime_mask: NDArray[bool] | None = None)
* Backwards-compatible: None → identical behavior to before
* When supplied, mask is applied at scoring time (ricci + target →
NaN outside mask), Ricci signal itself still computed on full
contiguous series (its rolling corr needs consecutive rows)
- tests/test_l2_regime.py — 7 new tests:
* shape + warmup on rolling_corr_regime
* high-ρ vs low-ρ synthetic discrimination
* argument validation (window too small, single symbol)
* mask NaN handling
* killtest rejects wrong mask shape
* trivial all-True mask matches unconditional (regression)
All 26 tests green. ruff + black + mypy --strict clean.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
GeoSync/research/microstructure/killtest.py
Line 459 in 8c31f7b
Report effective sample count for masked killtest runs
n_samples is always reported as features.n_rows, but with regime_mask many rows are excluded from IC/p-value computations, so the verdict metadata overstates the amount of data actually used. This can mislead analysis scripts or operators about statistical power; n_samples should reflect the finite scored rows (or an additional effective-sample field should be emitted).
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
Comment on lines
+380
to
+381
| ricci_panel = np.where(panel_mask, ricci_panel, np.nan) | ||
| target = np.where(panel_mask, target, np.nan) |
There was a problem hiding this comment.
When regime_mask is applied, rows outside the regime are converted to NaN and then _permutation_pvalue still shuffles across the full time axis; for sparse masks this makes most shuffled rows miss the finite target rows, shrinking effective trial sample size (often below _pooled_ic's minimum) and biasing permutation p-values toward non-significance. This can falsely KILL regime-conditional runs even when the observed IC is strong, so the null test should permute/compress only the active (finite) rows.
Useful? React with 👍 / 👎.
Two more diagnostic scripts for the regime filter, completing the generalization ladder: * scripts/l2_regime_walkforward_calibration.py Rolling 60-min calibration + 30-min evaluation. Slides across substrate; at each step derives q50/q75 thresholds from the calibration window, applies them to the next evaluation window. Result on collected substrate: uplift POSITIVE in only 1 of 7 steps (q50) / 1 of 5 steps (q75). HONEST LIMIT: the 50/50 split uplift (IC +0.12 → +0.24 OOS) does NOT survive production-style short-window rolling recalibration. Threshold needs longer calibration horizons to stabilize. * scripts/l2_regime_cross_session.py Cross-session OOS scaffold: takes --train-dir and --test-dir, derives quantile thresholds from the train session, applies to the test session, writes results/L2_REGIME_CROSS_SESSION.json. Runnable against the second 8h session currently being collected into data/binance_l2_perp_v2. Strongest form of OOS we can do without multi-day walk-forward. These land as diagnostics only. The regime MODULE (regime.py) and its integration (run_killtest regime_mask param) are the shippable artifacts. Scripts document the calibration surface honestly — including where the filter breaks under stricter recalibration. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
6 tasks
neuron7xLab
added a commit
that referenced
this pull request
Apr 18, 2026
Open question from PR #237: does Ricci cross-sectional signal polarity vary with UTC hour-of-day, or was Session-2's IC=-0.22 just sample noise? Binary test via new research/microstructure/diurnal.py folds three collected sessions into UTC-hour buckets and runs per-hour pooled Spearman IC + permutation test. VERDICT: SIGN_FLIP_CONFIRMED Per-hour IC (pooled across sessions; all buckets at permutation p<0.05 unless noted; n_rows after NaN-filter): hour IC p sessions regime 05Z n/a — S3 underpowered 06Z -0.0619 0.002 S3 Sat AM quiet 07Z -0.1194 0.002 S3 Sat AM quiet 08Z +0.0086 0.036 S1+S3 EU open weak 09Z +0.0191 0.002 S1+S3 EU open 10Z +0.1279 0.002 S1 EU active 11Z -0.0360 0.002 S1 EU midday lull 12Z +0.0556 0.002 S1 EU + pre-US 13Z +0.3554 0.002 S1 EU+US overlap PEAK 20Z +0.0008 0.954 S2 underpowered (n=7410) 21Z -0.0715 0.002 S2 US close / EU eve 22Z -0.1989 0.002 S2 EU evening Significance tally: 5 positive hours + 5 negative hours at p<0.05. Verdict: SIGN_FLIP_CONFIRMED (at least one bucket each side). Architectural implication: The regime-q75 filter landed in PR #236 conditions on volatility but does NOT capture the inversion driver. Saturday-morning 07Z shows IC=-0.12 at ordinary vol levels — a pure vol threshold does not separate this from the +0.13 of 10Z. Time-of-day is a load-bearing axis the current gate does not model. Next step (out of scope for this PR): diurnal-aware sign strategy as follow-up research module. Components: research/microstructure/diurnal.py (typed, 7 tests) utc_hour_of_row(start_ms, n_rows) -> NDArray[int64] compute_diurnal_profile(sessions, horizon_sec, min_rows_per_hour, perm_trials, pvalue_gate, seed) -> DiurnalProfile profile_to_json_dict(profile) -> dict session_start_ms_from_frames(frames) -> int scripts/run_l2_diurnal_profile.py --data-dir PATH (repeatable; one per session) --horizon-sec INT (default 180) --min-rows-per-hour INT (default 300) --perm-trials INT (default 500) --pvalue-gate FLOAT (default 0.05) --seed INT (default 42) --output PATH tests/test_l2_diurnal.py (7 tests): utc_hour_of_row monotone + wrap + negative rejection empty-sessions underpowered low-sample underpowered/stable multi-session hour merging JSON schema contract determinism under fixed seed Evidence: results/L2_DIURNAL_PROFILE.json (tracked; full per-hour breakdown) Quality gates: ruff + black + mypy --strict clean. Test regression: 42 → 49 (+7 diurnal). Numerical locks (PR #238 fixtures) unchanged: ic_test_q75 = 0.23638402111955653 breakeven_q75 = 0.4072465349699599 Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
neuron7xLab
added a commit
that referenced
this pull request
Apr 18, 2026
#240) Open question from PR #237: does Ricci cross-sectional signal polarity vary with UTC hour-of-day, or was Session-2's IC=-0.22 just sample noise? Binary test via new research/microstructure/diurnal.py folds three collected sessions into UTC-hour buckets and runs per-hour pooled Spearman IC + permutation test. VERDICT: SIGN_FLIP_CONFIRMED Per-hour IC (pooled across sessions; all buckets at permutation p<0.05 unless noted; n_rows after NaN-filter): hour IC p sessions regime 05Z n/a — S3 underpowered 06Z -0.0619 0.002 S3 Sat AM quiet 07Z -0.1194 0.002 S3 Sat AM quiet 08Z +0.0086 0.036 S1+S3 EU open weak 09Z +0.0191 0.002 S1+S3 EU open 10Z +0.1279 0.002 S1 EU active 11Z -0.0360 0.002 S1 EU midday lull 12Z +0.0556 0.002 S1 EU + pre-US 13Z +0.3554 0.002 S1 EU+US overlap PEAK 20Z +0.0008 0.954 S2 underpowered (n=7410) 21Z -0.0715 0.002 S2 US close / EU eve 22Z -0.1989 0.002 S2 EU evening Significance tally: 5 positive hours + 5 negative hours at p<0.05. Verdict: SIGN_FLIP_CONFIRMED (at least one bucket each side). Architectural implication: The regime-q75 filter landed in PR #236 conditions on volatility but does NOT capture the inversion driver. Saturday-morning 07Z shows IC=-0.12 at ordinary vol levels — a pure vol threshold does not separate this from the +0.13 of 10Z. Time-of-day is a load-bearing axis the current gate does not model. Next step (out of scope for this PR): diurnal-aware sign strategy as follow-up research module. Components: research/microstructure/diurnal.py (typed, 7 tests) utc_hour_of_row(start_ms, n_rows) -> NDArray[int64] compute_diurnal_profile(sessions, horizon_sec, min_rows_per_hour, perm_trials, pvalue_gate, seed) -> DiurnalProfile profile_to_json_dict(profile) -> dict session_start_ms_from_frames(frames) -> int scripts/run_l2_diurnal_profile.py --data-dir PATH (repeatable; one per session) --horizon-sec INT (default 180) --min-rows-per-hour INT (default 300) --perm-trials INT (default 500) --pvalue-gate FLOAT (default 0.05) --seed INT (default 42) --output PATH tests/test_l2_diurnal.py (7 tests): utc_hour_of_row monotone + wrap + negative rejection empty-sessions underpowered low-sample underpowered/stable multi-session hour merging JSON schema contract determinism under fixed seed Evidence: results/L2_DIURNAL_PROFILE.json (tracked; full per-hour breakdown) Quality gates: ruff + black + mypy --strict clean. Test regression: 42 → 49 (+7 diurnal). Numerical locks (PR #238 fixtures) unchanged: ic_test_q75 = 0.23638402111955653 breakeven_q75 = 0.4072465349699599 Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Recursive + cyclic + rolling-walk-forward analysis of the 5h14m collected L2 substrate revealed that the Ricci κ_min cross-sectional edge is intermittent, not uniform — some time blocks emit IC > +0.18, others invert to IC < -0.10. Full-window PROCEED averaged these.
This PR introduces the regime filter that rescues the edge:
2.03× IC uplift with threshold learned on first half generalizing cleanly to second half.
Architecture (AE-compliant — elimination over addition)
New public surface in
research/microstructure/regime.py:One optional parameter added to existing gate:
None→ identical behavior to before. No new dataclasses, no new CLI flags (yet — they'll land when the regime filter earns its place in the permanent gate).Analysis scripts (diagnostics, not production)
scripts/l2_killtest_recursive.py— depth-first bisection (full → halves → quarters → octiles) + cyclic K=8 blocks. Exposes regime structure.scripts/l2_regime_analysis.py— per-block regime features + IC rank correlation.scripts/l2_walk_forward.py— 56 rolling 40-min windows with 5-min step, Spearman feature→IC correlations.scripts/l2_regime_conditional.py— in-sample conditional gate at multiple quantile thresholds × window sizes.scripts/l2_regime_oos.py— true OOS, threshold on train, applied to test.Quality gates
What's explicitly NOT in this PR
--regimeyet — waiting for 8h fresh collection (currently running in background) to confirm threshold generalizes across a second session, not just across halves of one session.Test plan
l2_regime_oos.pyagainst FRESH substrate with threshold from ORIGINAL session — if uplift holds across sessions, regime filter is production-ready.🤖 Generated with Claude Code