Risk managers building portfolio volatility models often assume that a more complex ensemble will outperform a simpler one. The evidence from walk-forward testing on multi-asset portfolios does not consistently support that assumption.
What Boosting Does in This Context
Gradient boosting applied to risk factor data iteratively refines predictions by correcting residual errors. It can capture non-linear relationships between macroeconomic indicators and asset volatility that linear models miss. But those non-linear relationships are often artefacts of the training period rather than stable structural features of financial markets.
What Blending Actually Achieves
Blending combines predictions from several independently trained models using fixed or learned weights. A simple blend of a GARCH model, a random forest, and a regularised regression has shown more stable out-of-sample calibration than boosted ensembles alone across multiple market regimes studied in academic literature on volatility forecasting.
- Blending preserves the diversity of modelling assumptions, which is valuable when no single model dominates across all market conditions.
- Boosting can overfit to volatility clustering patterns that are period-specific.
- Blending weights can be updated more transparently, which matters for internal model governance processes.
- Boosted models require more infrastructure to retrain and version-control in production risk systems.
Where Boosting Earns Its Place
Boosting genuinely outperforms blending when the risk model needs to incorporate a large number of structured features simultaneously, such as combining position-level data with macroeconomic signals and sentiment indicators. For narrower volatility estimation tasks, a well-calibrated blend is harder to beat and easier to explain to a risk committee.