Volatility Forecaster

Start with EWMA (lambda = 0.94). It's simple: σ²_new = 0.94 × σ²_old + 0.06 × r²_yesterday. It captures volatility clustering and updates easily. Most spreadsheets can implement it.

Forecasting accuracy decreases with horizon. 1-5 day forecasts are reasonably accurate. 10-30 day forecasts are moderate. Beyond 30 days, forecasts converge toward long-run average with limited added value.

IV is a useful forecast but biased high due to the variance risk premium. It overestimates realized vol by 2-4% typically. It's best combined with statistical forecasts rather than used alone.

Size inversely to forecasted vol: Position = Target Risk / (Price × Forecast Vol × √Holding Period). If forecast vol doubles, halve your position to maintain the same dollar risk.

News comes in clusters (earnings seasons, economic events), traders react to volatility itself (hedging increases in high vol), and fear/uncertainty persists. High vol breeds more high vol until conditions calm.

Use specialized software: Python 'arch' package, R 'rugarch', or EViews. Provide return series (at least 250 observations). The software maximizes likelihood to estimate ω, α, β. Check α + β < 1 for stationarity.

Both capture leverage effects well. GJR-GARCH is more common and easier to interpret (γ is extra impact of negative returns). EGARCH models log-variance so never goes negative. Try both; compare AIC/BIC.

Calculate MAE, RMSE, QLIKE between forecasts and realized vol. Use Mincer-Zarnowitz regression (regress realized on forecast; want a=0, b=1). Compare your model to a simple benchmark like 20-day historical vol.

HAR-RV uses daily, weekly, and monthly lagged volatility to forecast. It's simple (just linear regression) but captures multiple horizons well. Use it when you need forecasts for different horizons or want a robust, easy-to-implement model.

Weight them: Combined = w × IV + (1-w) × Model. Start with w = 0.3-0.5. IV provides market information; model provides structure. The optimal weight can be estimated by minimizing forecast error on historical data.

Sum squared intraday returns: RV = Σr²_i. Use 5-minute returns (balance between efficiency and microstructure noise). Apply noise corrections (kernel estimators, pre-averaging) if using higher frequency. 5-minute RV with subsampling is robust.

SV if: pricing options (captures smile), need vol-of-vol parameter, research context. GARCH if: forecasting (similar accuracy, easier), production systems, limited computational resources. For pure forecasting, GARCH is usually sufficient.

Use features like lagged RV, returns, VIX, ATR, volume. Try Random Forest or Gradient Boosting for regression. Use LSTM if you suspect complex temporal patterns. Always use proper time-series cross-validation (no future leakage). Often, combining ML with GARCH forecasts as features works well.

Extract IVs at different maturities. Apply PCA to get level, slope, curvature factors. Model factors with VAR or state-space model. Alternatively, adapt Nelson-Siegel to volatility. For trading, monitor term structure shape for mean reversion signals.

Architecture: Data pipeline → Multiple models (GARCH, EWMA, HAR, ML) → Forecast combination → Output API. Re-estimate models weekly/monthly. Monitor forecast accuracy continuously. Alert on unusual forecasts. Document everything. Use ensemble averaging weighted by recent performance.