How machine learning actually reads order flow — and where it fails
A plain-language walkthrough of the features quantitative desks feed their models, and why most retail 'AI signals' are curve-fitted noise.
Models that trade well do not predict price. They predict short-horizon imbalance: whether the next marginal order is more likely to lift the offer or hit the bid.
The features that carry signal
Queue position, cancellation rates, trade-size clustering and the shape of the book two levels deep. Almost all of the predictive content lives in seconds, not days.
Where it breaks
Regime change. A model fitted through a low-volatility grind learns that mean reversion always pays — until the day it does not, and the position sizing built on that assumption is what does the damage.
How to evaluate any 'AI signal' you are sold
Ask for out-of-sample performance across at least one violent regime, net of fees and slippage. If the answer is a backtest curve with no drawdown chart, you have your answer.