believe · v133

A FUSION-DUAL quantitative strategy on E-mini S&P 500 futures. An ML tick ensemble, an XGBoost 5‑minute bar model, and an MBP-10 microstructure classifier, each voting through independent OCA brackets at the exchange. v133 adds an adaptive regime gate (2-day rolling P&L auto-toggle) and a session cut-down to RTH + ETH_EUROPE only, trimming backtest drawdown ~30% at the same profit tier. Purged walk-forward validated. Marketable-limit execution. No black boxes, no moving goalposts.

ES CME Futures

MBP-10 Order Book

0.826 Walk-Fwd AUC

v133 Production

What believe is

Three independent models. One instrument. One execution path.

believe trades only ES futures on CME. We do not diversify across instruments to paper over a weak model. We diversify across decision processes: three models, trained on different horizons and different views of the same tape, each firing its own qty=1 marketable-limit bracket. If any one model is the only voice in the room, it is the only voice that trades.

01 / Tick ML

Order-flow sequence model

An LSTM-style sequence network consumes the tick stream — signed trade size, local imbalance, VPIN, Kyle’s λ, microprice drift, realised-volatility terms — and emits a short-horizon directional probability. Labels come from a triple-barrier scheme with embargo.

02 / XGB 5m

Structural bar model

Gradient-boosted trees on 52 bar-level features: trend strength, session context, range dynamics, volatility regime, relative position in the day. One decision per closed 5-minute bar. Walk-forward out-of-sample validated, no touching of the test fold until purge and embargo have elapsed.

03 / F2_dom

MBP-10 microstructure

A LightGBM classifier on 17 live order-book features — book imbalance, top-of-book ratio, top-three imbalance, bid and ask gradients across ten levels, rolling imbalance std, midprice momentum. Trained on 1.45M labelled samples. 0.826 ± 0.015 AUC under purged K-fold.

Why this, and not something else

Our edge is in the plumbing, not the forecast.

Everyone has a model. The difference is what you do between the model and the exchange. believe is designed around four choices that most retail quant stacks quietly get wrong.

Why ES

E-mini S&P 500 is the deepest, most liquid equity-index futures contract in the world. Tight spreads, minimal size impact at qty=1, round-the-clock price discovery. It rewards microstructure honesty and punishes overfitting — the right crucible for a research-grade system.

Why microstructure

Price is the last thing to move. The order book is the cause, the tape is the effect. F2_dom reads 10 levels of bids and asks around the clock. Structural imbalance, queue gradients, top-of-book pressure — none of it shows up in a 5-minute candle. All of it shows up before the candle closes.

Why purged walk-forward

Standard K-fold CV leaks the future into the past through overlapping labels and autocorrelated features. We use López de Prado purged K-fold with embargo, triple-barrier labels, and isotonic calibration on held-out folds. The reported AUC is what a deployed model would have seen, not a curve-fit fantasy.

Why marketable limit

Market orders quietly cost you a tick every fill. Over thousands of trades, that tick is the difference between a real edge and an imagined one. believe enters with marketable limit orders — aggressive enough to fill, disciplined enough to refuse worse prices. Every order attaches a server-side OCA bracket before acknowledgement.

Walk-forward, not wishful

How we measure ourselves.

We do not publish daily P&L on a marketing site. What you see below is the kind of statistic that belongs in a risk review: classification quality of the microstructure model on purged, embargoed folds; hit rate and profit factor of the full stack against 79 days of real captured tick data (RTH only, adaptive regime gate on).

0.826 F2_dom AUC

76.1% Stack Hit Rate

1.66 Profit Factor

−30% Max DD vs v131

F2_dom_v133 classification metric is purged K-fold (5 folds, 10-minute embargo) on 1.45M labelled samples, 2026-01-27 to 2026-04-15. Stack hit rate and profit factor are from a 79-day backtest (Jan 29 – Apr 17 2026, RTH only) on real captured tick and MBP-10 data, qty=1, with commission and a ±1-2 tick stop-slippage model, adaptive regime gate on. Max-DD reduction is v133 (adaptive gate) vs v131 (no gate) on the same tape. Past results do not guarantee future performance.