Predicting Trade Signals from Level-2 Limit Order Book Data Using Time Series Models and Profit Simulation

Today's electronic markets execute millions of orders per second, and slight improvements in execution efficiency can translate into significant profits. In this paper, we introduce a new hybrid forecast model that integrates TCNs with ARIMAbased residual correction for making short-term trading signal prediction using Level-2 Limit Order Book (LOB) data. In contrast to other methods that depend on either deep learning or statistical models, our method can capture nonlinear temporal features and structure residual models, resulting in enhanced prediction performance. We systematically compare conventional time series models (ARIMA, SARIMAX) and deep learning architectures (LSTM, GRU, TCN) via a unified pipeline that involves data preprocessing, feature engineering, forecasting, and profit calculation. Back-testing and Monte Carlo simulation indicate that the method specializes in an 8224 % cumulative return with a 10.9 Sharpe ratio. The results accentuate that of hybrid modelling in HFT, and emphasize the significance of feature design, robustness check, and real-world constraints in developing an algorithmic strategy.