Multi scale LSTM model construction and NSGA-II optimization method for dynamic threshold of high frequency trading signal

Forecasting high-frequency trading signals is complicated by microstructure noise, heterogeneous temporal patterns, changing market states, and strict low-latency requirements. This paper constructs a multi-scale LSTM and dynamic threshold attention framework, and adopts a candidate-level multi-objective configuration screening method under deployment constraints. Level 3 order data are transformed into microstructure features that describe large-order flow, order-flow imbalance, order-book pressure, liquidity pressure, high-frequency volatility, and cross-market cointegration residuals. DB4 wavelet decomposition separates high-, medium-, and low-frequency components, which are encoded by independent LSTM branches. Attention sensitivity is conditioned on volatility and liquidity, while NSGA-II searches the configuration space using predictive performance, maximum drawdown, and inference latency as joint objectives. The empirical analysis uses 30 highly active CSI 300 constituents over 240 trading days in 2022. The base multi-scale dynamic-threshold architecture reports an AUC of 0.861, a Sharpe ratio of 5.6, a maximum drawdown of 11.5, and an inference latency of 15.9 μs. The final NSGA-II-selected configuration improves these point estimates to an AUC of 0.867, a Sharpe ratio of 5.8, a maximum drawdown of 10.7, and an inference latency of 12.8 μs. The one-minute forecasting horizon and the 3 to 5 percent volatility regime provide the strongest reported performance for the base architecture. Rolling-window, sector, feature, and event analyses are used to assess robustness without treating descriptive associations as causal evidence.