Deep Learning on High-Frequency Order Flow: Predicting Trade Direction in Futures Markets

Abstract

In futures markets, a continuous stream of limit order book quote updates and trade events often encode predictive signals about price movements and order flow. Extracting these signals reliably is challenging: the data is noisy, non-stationary, and structurally different from the fixed-interval time series that most standard machine learning pipelines assume. This thesis explores how the choice of data representation and model architecture shapes the predictive signal available at different time horizons. Using high-frequency futures market data from the Chicago Mercantile Exchange (CME), we study two complementary representations: engineered features sampled on a fixed time grid, and raw event sequences where each update is treated as a discrete token. We find that meaningful directional autocorrelation exists in the event stream, but is concentrated at very short horizons—essentially the next one or two trades—and decays rapidly as the prediction window grows. Sequence models trained on events rather than time bars capture this structure more effectively, with a decoder-only Transformer achieving 78.9% trade-direction accuracy against a 67.7% persistence baseline. At longer horizons, where the autocorrelation has dissipated, even carefully calibrated models extract only marginal signal above baseline, underscoring a fundamental limit to price prediction in futures markets.