Every change to the engine, in the open.
Most trading services hide their evolution. We publish every phase, including the ones we threw out. 12 shipped, 5 reverted. The current live baseline is Phase 37 — shipped May 17, 2026.
- Phase 37May 17, 2026+163.33% OOS @ 10 bps · Sharpe 3.31 · MaxDD −13.6%Sleeve-class quota (carry 10% / trend 90%)
Re-enabled basis/cash-and-carry sleeves with a hard 10% capital cap so they can't Sharpe-trap the trend allocation. Strict domination over Phase 33: Sharpe up and MaxDD down at every slippage level (0/5/10/20 bps), PnL flat. Current live baseline.
- Phase 33May 2026+163.07% OOS @ 10 bps · Sharpe 3.21 · MaxDD −15.3%James-Stein Bayesian weight shrinkage
Top-N weights shrunk toward equal-weight prior with λ=0.25. Reduces overfit to in-sample Sharpe; tighter drawdowns with unchanged returns. Superseded by Phase 37.
- Phase 24May 2026+283.19% OOS @ 10 bps · Sharpe 2.07Greedy sleeve decorrelation
Allocator now walks the Sharpe-ranked sleeve list and only adds a sleeve if its train-window correlation with every already-picked sleeve is below a threshold. Adds +56pp over Phase 23 with the same drawdown floor.
- Phase 23Apr 2026+226.62% OOS @ 10 bps · Sharpe 1.95Regime-aware Sharpe-weighted top-5
BTC trend classifier overlay shifts weights toward short-vol and hedge sleeves during regime-off periods. Validated by splitting OOS windows into BTC-up vs BTC-down buckets.
- Phase 22Apr 2026−14pp return for +0.4pp DDWF persistence filter (reverted)
Required each sleeve to appear in the previous top-2N pool. Paid too much performance for too little drawdown improvement. Reverted.
- Phase 21Mar 2026+82.59% @ 20 bps · Sharpe 1.62Sharpe-weighted allocator
Replaced equal-weight ensemble with weights proportional to max(0, train Sharpe). Eliminates noisy-loser drag.
- Phase 20Mar 2026More OOS observationsWalk-forward history extended to 730 days
Doubled price-history depth for the walk-forward harness while keeping regular backtests at 365d. Unlocks 270/365d train windows.
- Phase 19Feb 2026DiversificationLow-turnover and commodity sleeves
Triple-EMA, monthly momentum, slow Donchian and gold/silver sleeves added to the catalog. Lowered overall portfolio turnover and added a non-crypto correlation hedge.
- Phase 18Feb 2026Robustness verificationSlippage stress test
Walk-forward runs now evaluate at 0 / 5 / 10 / 20 bps round-trip cost. Any new sleeve that doesn't survive 10 bps is rejected as curve-fit.
- Phase 17Jan 2026OOS validationWalk-forward harness
Rolling 180-day train / 30-day test windows. Catalog re-ranked by train-window Sharpe each step, OOS returns concatenated.
- Phase 15Dec 2025Uncorrelated down-regime PnLShort-only crash sleeve
Vol-spike short sleeve activated only when BTC regime is off. Adds an uncorrelated PnL stream during down markets.
- Phase 13Nov 2025PnL fell to 64.85%Cluster cap (reverted)
Capped concurrent positions in the BTC/ETH/SOL/BNB/AVAX cluster. Cut PnL AND worsened max drawdown. Reverted.
- Phase 12Nov 202575.22% → 74.25%Second pyramid add (reverted)
Tried a second add at +2R. Phase 11 single-add proved to be a tight local optimum; reverted.
- Phase 11Oct 2025~+75.22% standalone ceilingPyramid into winners
Pyramid trigger at +1R with original stop moved to breakeven. Established the standalone-strategy ceiling that the meta-portfolio later breaks.
- Phase 8Sep 2025PnL regressed to 28%Combined trade filters (reverted)
Weekend-skip combined with low-ATR percentile gate killed too many profitable entries. Reverted.
- Phase 6Aug 2025Worse PnL vs Phase 4Funding gate + vol-target + heat cap (reverted)
Three tightening controls layered together. Looked smart on paper, regressed in walk-forward. Reverted.
- Phase 4Jul 2025First profitable in-sample configBaseline breakout
Donchian-20 breakout with BTC regime filter, ATR trailing stop and run-winners exit. The starting point that every later phase builds on.