Open-source implementation of HAR-CEV for zero‑shot realized variance forecasting
The repository implements the HAR‑CEV model for forecasting daily realized volatility across economic regimes, offering training scripts, baseline comparisons, and VaR backtesting. It works with intraday data, converts it to realized volatility, and provides a full evaluation pipeline for researchers and quantitative analysts. By combining HAR components with a context‑dependent CEV transition, it outperforms standard econometric baselines and gives reproducible results. The code is organized, uses PyTorch, and includes utilities for data validation and metric computation.
View on GitHub →finxlab/HAR-CEV