Machine Learning Forecasting of Cryptocurrency Realized Volatility Using High-Frequency Market

I will study how high-frequency trading activity, transaction structure, market noise and recent price variation affect short-term cryptocurrency volatility. The dataset covers Bitcoin, Ethereum, BNB, Dogecoin, Solana and XRP on the Binance spot market from April 2023 to April 2026. Forecasts will be produced every five minutes for realized volatility over the following hour. I will construct predictors from past volatility, trade counts, turnover, trade-size distributions, buy–sell imbalance, noise-related price measures and selected financial-market indicators. The forecasting exercise will compare linear models with Random Forest, XGBoost and LSTM. Each model will be evaluated using a 365-day rolling training window, chronological validation and regular refitting, so that every forecast uses only information available at that time. Accuracy will be compared across currencies, information sets and market conditions using RMSE, MAE and QLIKE, together with forecast-comparison tests and robustness checks. The main question is whether transaction and noise-related information improves forecasts beyond historical volatility, and whether nonlinear models use this information more effectively than conventional volatility models. The experiment will retain model settings, predictions and evaluation results for each rolling period, allowing the complete analysis to be reproduced.