Falsification-first quant research. One supported edge (its confidence interval excludes zero; not yet independently confirmed). Every other hypothesis here was rejected or not promoted, and each row gives the mechanism.
Research practice: decision rules pre-registered for most studies (not for the TSMOM core, which predates that protocol and is now under a prospective test, nor for the crash-defense and breakout overlays) · BH-FDR within each registered family · look-ahead checked with truncation-invariance tests (two engine leaks were found and fixed in quant-backtest-framework; its re-run on the corrected engine held the verdict) · negatives reported with the same prominence as positives.
| Hypothesis | Repo | Verdict | Mechanism | Key stat |
|---|---|---|---|---|
| Multi-asset TSMOM core (17 ETFs, vol-targeted) | multi-asset-tsmom-research | SUPPORTED: CI excludes 0; not independently confirmed; prospective test live since 2026-09-13 | diversified trend premium; positive in the 2008 and 2020 crisis windows (raw return, not benchmark-adjusted) | net Sharpe (2 bps): 0.75 95% CI low (2 bps): 0.29 95% CI high (2 bps): 1.23 net Sharpe, sensitivity: net of 3-month T-bill, whole book: 0.61 95% CI, sensitivity: net of 3-month T-bill, whole book: [0.15, 1.09] net Sharpe (5 bps): 0.70† 95% CI (5 bps): [0.24, 1.18]† GFC 2008 window return: +11.6% COVID 2020 window return: +7.3% prospective test (C-A) live since: 2026-09-13† |
| MTF SMC price action replicates across instruments | quant-backtest-framework | FALSIFIED: re-run on the engine corrected on 2026-09-27 (run at commit 631f33d, code_hash 07f155612e0c70dc); the pre-registered rule (docs/SPEC_multi_instrument.md §7) gives the same verdict as before the fix | No configuration is positive and significant on even one of the five instruments, 0/210 cells survive the cross-instrument BH-FDR, the L1 walk-forward is negative out of sample, and cross-instrument return correlation leaves an effective 3.45/5 independent instruments. | config x instrument cells surviving cross-instrument BH-FDR: 0/210 configs positive-and-significant on >= 2 instruments: 0/42 L1 walk-forward pooled OOS E[R]: −0.329 R L1 walk-forward window-block 95% CI: [−0.416, −0.228] L1 walk-forward calendar OOS periods with E[R] < 0: 11/12 (one-sided p = 0.0032) effective number of independent instruments: 3.45 |
| Crash-defense overlay (de-gross when systemic risk spikes) | multi-asset-tsmom-research | not_promoted: rejected at Phase 0; premise not supported (descriptive comparison, no inference) | the causal systemic-risk trigger (portfolio-vol percentile) reads 0.94–0.95 in the 2008 and 2020 crisis windows, where the core made money, and 0.49 in its actual drawdown months | trigger percentile, GFC 2008: 0.95† trigger percentile, COVID 2020: 0.94† trigger percentile, drawdown months: 0.49† |
| Vol-compression breakout overlay | multi-asset-tsmom-research | not_promoted: rejected at Phase 1B; premise not supported (descriptive comparison, no inference) | compression is followed by more volatility (1.31× vs an unconditional 1.10×) but not by direction (efficiency-ratio change ≈ 0) | vol expansion after compression: 1.31× vol expansion, unconditional: 1.10× efficiency-ratio change: +0.004 |
| Seasonality / calendar-effects overlay | multi-asset-tsmom-research | not_promoted: rejected at premise, 0/18 | no calendar cell survives BH-FDR (q = 0.10) across the pre-registered 18-cell family; 1 of 18 raw p < 0.05, as expected under noise | cells passing BH-FDR: 0/18 smallest BH-adjusted p: 0.48 cells with raw p < 0.05: 1/18 |
| Yield-curve slope as a macro-regime conditioner | multi-asset-tsmom-research | not_promoted: rejected at premise, 0/6 | no cell significant (BH-FDR p 0.60–0.67); dropping the 2022–24 episode keeps 26–68% of the point estimate, sign preserved in 6/6 cells | cells confirmed: 0/6 BH-FDR adjusted p: 0.60–0.67 share of the effect kept without 2022–24: 26–68% cells keeping their sign without 2022–24: 6/6 episodes in the tested flat state: 12–13 2022–24 share of flat-state days: 19–21% |
| XSMOM adds an independent edge | multi-asset-tsmom-research | falsified: 0/5 universes (BH-FDR at α = 0.05 + walk-forward + 3/6/9/12 sign) | +0.42 corr with TSMOM; decomposition cannot separate the terms (term1 and lead-lag CIs contain 0) | universes confirmed: 0/5 corr with TSMOM (17 ETFs): +0.42 term1 (own-autocorrelation) CI contains 0: 5/5 term2 (lead-lag) CI contains 0: 5/5 |
| Order-flow footprint signals (4 tested; 2 data-blocked, disclosed) | orderflow-research-engine | null (H1, H2); underpowered (H3, H6): the pre-registration's outcome classes: H1 and H2 fail BH-FDR; H3 and H6 fall below the 300-event gate; H4 and H5 are data-blocked and untested | best cell ~2 bp gross vs 18 bp materiality bar (1.5× ~12 bp round-trip cost); no OOS return statistic computed or reported | BH-FDR significant cells (q = 0.10): 0/20 smallest BH-adjusted q: 0.109 best cell gross mean (H1, 1h): 2.1 bp materiality bar (gate 3): 18 bp round-trip cost: 12 bp day-cluster bootstrap reps (20 primary cells): 2,000,000† H3 in-sample events: 62 H6 in-sample events: 286 |
| Spot-MFI momentum predicts perp returns | spot-mfi-btc-perp-research | FALSIFIED | non-overlapping IC ≈ 0 before any strategy (full-sample IC, OOS included); the full-sample grid peak collapsed in the walk-forward to an OOS Sharpe below both buy-and-hold benchmarks | BH-FDR survivors (M1 grid, full-sample): 0/42† full-sample peak Sharpe (M1 grid): 1.07† walk-forward OOS net Sharpe: 0.29† B&H over the OOS span (perp price ex-funding / registered: perp net of funding): 0.46 / 0.31† |
| MFI × funding divergence (Variant A) | spot-mfi-btc-perp-research | INCONCLUSIVE, leaning FALSIFIED | Edge IC ≈ 0 and weaker than raw MFI; walk-forward OOS Sharpe beat buy-and-hold (perp price ex-funding) on OOS windows reused from the base study, then failed the pre-registered bootstrap-CI and permutation gates (the base study's M2 also beat buy-and-hold OOS and was never tested) | walk-forward OOS net Sharpe: 0.68† B&H (perp price ex-funding): 0.46† OOS Sharpe 95% block-bootstrap CI: [−0.19, 1.62]† permutation p (gross returns): 0.10† |
| Cross-sectional commodity carry (18 CME futures) | commodity-carry-research | NOT PROMOTED: fails 4 of the 5 pre-registered promotion gates (only net Sharpe > 0 under the 2x cost table passes); corrected re-run of 2026-09-28, reports/ADDENDUM_2026-09-27.md §10 | no confirmable edge net of costs; none of H1's 9 registered variants reaches the 0.30 gate; the long leg's gains outweigh the short leg's losses (exploratory, no CI) | net Sharpe: 0.059 95% CI low: −0.41 95% CI high: 0.52 registered variants reaching the 0.30 gate: 0/9 |
| Time-series commodity carry (own-basis sign) | commodity-carry-research | NOT PROMOTED: closed at the sign-only premise gate (Sec 7): the premise point estimate is negative, so no backtest was run; corrected re-run of 2026-09-28 (the published run passed the gate and failed all 5 promotion gates) | sign of own carry has no detectable predictive slope (premise coefficient negative, t = -0.07); closed before any backtest | premise coefficient: −0.000143 premise t (clustered): −0.07 |
Generated by scripts/render_readme.py from each repo's results/headline.json; every key stat links to the file that holds it, at the commit that headline file names. Unmarked stats were recomputed from that file. † repo-reported, not reproduced (the inputs are licensed or no longer retrievable).
Other work: semiconductor-yield-screening: Semiconductor yield detractor screening (UCI SECOM, 590 inline sensors). MOSTLY AN ARTIFACT OF TIME: post-hoc holdout replication is weak and the predictive baseline is at no skill. The failure rate drifts over the collection window and many train-screen survivors trend with time; once the null keeps time order the screen only narrowly beats chance, and most associations do not transfer to the holdout (screen p, time order preserving null: 0.032; features replicating on holdout: 2 of 32).
In progress — NQ intraday programme: nq-intraday-research publishes one preregister-then-test programme on Nasdaq-100 futures as three projects with full history: the framework ITSF (opening-drive continuation; parked 2026-09-16, no governed result), R1 scheduled-release (CPI/NFP) diffusion continuation (one sealed, once-only test; closed, unresolved: the stored interval [−13.92, −4.11] NQ points per trade fails the sealed precision rule, half-width 4.91 > 3.99), and R2 leveraged-ETF daily-reset rebalancing pressure (pre-S1: no hypothesis stated until a point-in-time fund-size source is verified; a forward collector has captured the five funds' published AUM daily since 2026-09-22). A read-only multi-agent rigour audit of R1 and R2 (2026-10-01) found no Critical or High defect; its 25 record-level findings and their corrections are in the repository. No promoted strategy.
Two statistical traps, caught by the tests:
- A multiple-testing null. In the seasonality overlay, 1 of 18 calendar cells had raw p < 0.05, about what noise alone produces, and none survives BH-FDR.
- A fake winner. MFI × funding divergence beat buy-and-hold out of sample (0.68 vs 0.46), but on out-of-sample windows the base study had already used, and the base study's M2 model had also beaten buy-and-hold there (0.72) without ever being tested. It then failed its pre-registered bootstrap-CI and permutation gates (CI [−0.19, 1.62], p = 0.10).
Data defects caught: an exchange-side gap in Binance aggTrades on 2022-09-06 (quarantined; the window file is not yet published) and a same-ID / revised-quantity aggTrades mismatch (regression-tested), both in orderflow-research-engine; and a gold-calibrated absolute-price slippage constant that produced about −25 R per trade on EURUSD (repo-reported; the buggy run was not preserved), now regression-tested in quant-backtest-framework.
Audit record: an independent engineering audit of every repository above, run 2026-09-27, is published in audit/: findings, fixes, what each verdict now rests on, and the two re-runs on the corrected code.
Currently: MSc Data Science (Monash), targeting quantitative research roles. (https://www.linkedin.com/in/aaron-lau-b65b22224/)
