Long-Term Memory Modeling of Financial Volatility Based on Transformer Architecture

Long-horizon volatility forecasts support risk control, hedging, and planning. This study builds a long-memory model named TransVol and adds a LightGBM step to adjust short-term changes. The model uses daily CSI 300 data and is trained with a rolling setup to avoid future information. Performance is measured with mean absolute percentage error. Results show that TransVol-LightGBM lowers MAPE by 8.2% compared with an LSTM-LightGBM baseline when the horizon is longer than 30 days. These findings suggest that long-range attention helps capture slow shifts in volatility, while the correction step improves near-term accuracy. The model can help with monthly portfolio work and risk planning. A main limitation is that only daily data are used; adding order-book or macro inputs may help in periods of fast market change.