Volatility Forecaster

For most traders, checking once daily (after market close) is sufficient. Volatility regimes change gradually over days, not minutes. Set up alerts for significant changes so you don't need to constantly monitor. Intraday traders may want more frequent updates, but daily forecasts are the standard.

If your forecasts consistently beat a random walk (using yesterday's volatility as the forecast), you're doing well. Perfect forecasting is impossible. RMSE should be significantly lower than the benchmark, and directional accuracy above 55-60% is meaningful. Focus on whether forecasts improve your trading decisions rather than exact accuracy metrics.

Both have value. Historical volatility tells you what actually happened; implied volatility tells you what the market expects. For options trading, comparing the two is powerful - if IV is much higher than your historical estimate, options may be overpriced. For position sizing and stops, a blend of both often works best.

Volatility can behave unexpectedly in the short term. Sometimes it rises even when the market is stable (anticipating upcoming events). Sometimes it falls during slow declines (no panic, just drift). The typical inverse relationship between volatility and the STI is a general tendency, not a rule. Because Singapore has no flagship volatility index, watch STI realized volatility, STI option IV, and the global CBOE VIX together, and focus on the overall level and trend rather than daily fluctuations.

Short-term forecasts (1-5 days) are generally most accurate. GARCH and similar models provide useful forecasts up to about 20-30 days. Beyond that, forecasts converge to long-term average volatility and add limited value over just using the historical average. For longer horizons, focus on term structure from options rather than point forecasts.

Start with GJR-GARCH (captures asymmetry) with Student-t errors (captures fat tails) - this is a good default for equities. Compare models using information criteria (AIC, BIC) and out-of-sample forecast accuracy. The 'best' model statistically may not be best for your trading application, so also evaluate based on how forecasts impact your trading performance.

Weekly or monthly refitting is common. More frequent refitting captures recent dynamics but adds noise. If market structure changes significantly (regime shift, crisis), immediate refitting may help. Monitor whether recent forecast accuracy is degrading - if so, refit sooner. Balance stability (less frequent) with adaptability (more frequent).

Your forecast uses historical data and models; implied volatility comes from current option prices (STI options, where liquid) and the global CBOE VIX. Differences are normal and potentially valuable: 1) your forecast may lag recent information that the market has already priced in, 2) implied volatility includes a volatility risk premium (typically a few points above realized), 3) implied measures are usually for a ~30-day horizon while your forecast may target a different horizon. Use the comparison as a trading signal.

For known events: 1) implied volatility already prices in expected event volatility, so lean on option IV and term structure where available, 2) note that MAS moves the SGD via the S$NEER band rather than a policy rate, so the volatility shows up first in the currency and in rate-sensitive sectors (banks, REITs), 3) the US Fed is often a larger STI driver than domestic events, given Singapore's global linkages, 4) post-event, expect a volatility crush if the outcome lands within expectations. The ensemble approach helps by including implied vol, which is event-aware.

Intraday data can improve forecasts through 'realized volatility' measures (sum of squared 5-minute returns). However, this adds complexity and data requirements. For most users, daily data with proper methods (Yang-Zhang) is sufficient. Consider intraday approaches if you need very short-term forecasts or are implementing sophisticated institutional-grade systems.

Key components: 1) Reliable data pipeline with quality checks, 2) Model layer with GARCH, IV processing, optional ML, 3) Ensemble combiner with tracked weights, 4) Output API serving forecasts to trading systems, 5) Monitoring for accuracy degradation and alerts, 6) Scheduled reestimation with version control. Consider redundancy, fallback forecasts, and audit trails. Start simple and add complexity based on demonstrated value.

ML may add value when: 1) You have informative features beyond price (volume, options data, macro), 2) There are non-linear patterns GARCH doesn't capture, 3) You have sufficient data for training without overfitting, 4) You implement proper walk-forward validation. Often ML adds modest improvement over GARCH - sometimes none. Hybrid approaches (ML on GARCH residuals, or GARCH as ML feature) often work best. Always validate that complexity translates to better trading outcomes.

Portfolio volatility depends on asset volatilities AND correlations. Approaches: 1) Forecast individual volatilities + use DCC-GARCH for correlations, combine, 2) Directly model portfolio returns with univariate GARCH (simpler but loses granularity), 3) Factor approach - decompose into factors, forecast factor volatilities. Consider that correlations increase in stress (diversification fails). For risk management, use stress-tested correlations alongside forecasted volatilities.

Surface forecasting is harder to validate than ATM vol. Approaches: 1) Track accuracy of ATM level, term structure slope, and skew separately, 2) Evaluate economic value - do trades based on surface forecasts (skew trades, calendar spreads) generate P&L? 3) Use factor decomposition - forecast factors, validate each factor's accuracy, 4) Compare to benchmarks like random walk on each surface point or simple parametric extrapolation. Accept that surface forecasting is more uncertain than level forecasting.

Common pitfalls: 1) Overfitting - complex models that don't generalize, 2) Look-ahead bias - using future information in backtests, 3) Survivorship bias - only testing on data that's available now, 4) Ignoring transaction costs - theoretical edge lost to costs, 5) Not handling regime changes - model trained on calm period fails in crisis, 6) Over-reliance on IV - contains risk premium and may not predict realized vol, 7) Not tracking accuracy - continuing to use degraded forecasts. Mitigate with rigorous validation, continuous monitoring, and healthy skepticism.