Real‐Time Prediction System for High‐Frequency Trading Volatility Based on Integrated Learning

Volatility forecasting is important for intraday risk control and trade execution in fast-moving financial markets. This study builds a minute-level forecasting framework that combines LightGBM, XGBoost, and CatBoost with an adaptive sliding-window design. The approach is tested on one year of NASDAQ data from 150 active stocks. Realized volatility is computed from intraminute returns. Time-ordered evaluation is used to avoid the use of future information. The ensemble improves forecasting accuracy by 5.9% compared with the strongest single model. Median prediction time remains below 80 ms, which is suitable for real-time use. These results show that combining several boosting models and adjusting the window based on recent error helps follow short-term changes in market behavior. The framework can support intraday risk checks and execution decisions.