TradingBot

A raw opening-range breakout on US30 is barely profitable. Backtested honestly with a realistic ECN spread it sits near break-even, around a 35% win rate at two-to-one. There is an edge in there, buried in noise.

So the task is not predicting the market. It is: given everything observable at the moment a breakout fires, estimate the probability that this specific trade reaches +2R before −1R, and take only the trades whose calibrated probability clears a learned threshold. Framed that way it becomes measurable — the model is judged on calibration and expected R per trade, not on accuracy, and the decision to deploy comes from out-of-sample expectancy rather than a good-looking backtest.

The training set comes from replaying the strategy bar by bar across eight years of five-minute data and labelling each trade by what actually happened. Every trade is described by 67 features, all computed causally from information available at entry: the geometry of the setup, multi-timeframe context joined by an as-of merge against the last closed bar on each timeframe, and session and calendar position.

Features that looked predictive but carried no real information — raw price levels, CFD tick volume — were dropped after analysis rather than kept because they improved a number.

Most machine learning applied to trading fails because the evaluation leaks, not because the model is weak. The pipeline is built around that:

Time-ordered splits, never random

— train, calibration and test are contiguous in time, so the model is always tested on its own future.

Purged cross-validation with a gap and an embargo

so a trade whose daily-timeframe lookback overlaps the next segment cannot leak its label across the boundary.

The threshold is chosen on calibration and reported on test

. Picking the operating point on the same data you report is the most common way a backtest gets inflated.

A feature-parity test as the deploy gate

— a golden-vector test asserts the live feature builder produces the same vectors as the training pipeline, to float tolerance. It caught a real look-ahead bug where the two disagreed.

The filter is an ensemble of gradient-boosted trees, XGBoost and LightGBM, each tuned by Optuna over 200 trials optimising the Brier score rather than accuracy — because the strategy depends on the probabilities being honest, not merely well-ranked. Raw scores then pass through isotonic regression fitted on the held-out calibration set, so a 0.6 really does mean roughly a 60% win rate, and that claim is checked with reliability diagrams and log loss.

SHAP values rank what the model is leaning on, and it is retrained on the top 70% of features to confirm the edge is not resting on a few fragile inputs. Risk of ruin is modelled separately: bootstrap intervals on win rate and average R, and a Monte-Carlo trade shuffle to estimate the drawdown distribution and the chance of breaching a prop-firm limit at a given risk per trade.

A filter that merely takes fewer trades can look good by luck. So the pipeline compares the average R of ML-selected trades against taking every rule-filtered signal, and bootstraps the difference over 5,000 resamples. The model ships only if the lower bound of that interval stays above zero.

That test rejected a variant I wanted to keep. A one-to-one configuration produced a much prettier 63% win rate — but its confidence interval for expectancy included zero and one walk-forward fold went negative. The two-to-one configuration, with a worse-looking win rate, held up in every fold. The prettier number lost.

At the New York open the bot builds the fifteen-minute range, takes the first five-minute close beyond it, filters weak setups by rule, then scores the survivor and trades only above threshold. Stop at the volume-profile point of control, target at 2R, one re-entry allowed after a stop, flat by the end of the session.

Risk is fixed at 1% per trade, halved after two consecutive losses, behind a daily-loss breaker and a maximum-drawdown breaker. Breaker state is persisted, so a restart in the middle of a bad session does not hand the bot a clean slate it has not earned.