The graveyard
1,000+ tests. 2 shipped.
Across 17 ML paradigms, 50+ named sprints, and roughly a thousand hyperparameter configs since 2024, two strategies cleared every audit gate (R86b · R99b). The rest live here — categories, then 21 recent rejections.
By research direction
Each row is a paradigm we systematically tested — full post-mortems live in our internal experiments log.
ML overlay (meta-learning)
Add a model on top of existing strategies — pick winners, scale by confidence, skip losers. All five attempts hurt Sharpe via diversification loss.
Direct ML direction prediction
Predict next-bar direction with LightGBM / XGBoost on price + cross-asset features. AUC ceiling 0.50-0.51 — no extractable signal at hourly resolution.
Position sizing & filter schemes
Adjust position size or skip trades by volatility, regime, or time-of-day. Vol-target had 60 sub-configs; one passed but overfit a single hyperparameter.
Statistical arbitrage
Trade the spread between cointegrated assets reverting to long-run mean. XAU/XAG not cointegrated (p=0.26), half-life 142 days — untradeable.
Alternative data
Non-price signals (news, sentiment, social) for direction. GDELT raw + smart variants: no |IC| > 0.05 stable in both IS and OOS.
Volatility prediction
Predict next-period vol and scale exposure. Vol IS predictable (Spearman 0.17) — but unexploitable on retail XAU spot without options access.
HMM regime classifier
3-to-5-state Gaussian Hidden Markov Models to detect trend/chop/crisis. Train +3.8 → OOS +0.18 collapse. Overfits on 75 train trades.
COT sentiment filter
CFTC Commitment-of-Traders speculator positioning as entry gate. Pre-2025 pattern broke post-2025 — random null P=49.6%.
Sprint D LightGBM grids
Direction prediction on raw bars across 35 hyperparameter configs. Ceiling -0.5 Sharpe. Inversely-predictive in some folds.
Liquidity-grab / microstructure
Trade liquidity hunts at swing highs/lows. Sharpe -0.97 at hourly resolution. Needs Level-2 / tick data — retail-impossible.
Event-window search
Brute-force test every economic event × horizon combination. ~1000 tests with strict Bonferroni; only 42-Day Bill Auction survived (t=4.74).
21 recent sprint-level rejections
Detailed rejections from the 2026 sprint cycles, with the audit reason that killed each one.
Killed sprints stay published forever — both as a discipline forcing function and a public record of what we tried.