Before we settled on the current architecture for Metafide's vol forecasting pipeline, we ran a series of systematic comparisons across ML model families on held-out walk-forward data. The results were clarifying, not because every model performed well, but because the failure modes were instructive. This post is a technical account of what we tried, what degraded on out-of-sample data, and what held up well enough to include in production.
The framing matters: we are not trying to predict future realized vol with maximal accuracy for its own sake. We are trying to forecast the vol surface forward in a way that is useful for institutional risk management. That means the evaluation criterion is not just mean squared error on point forecasts. It includes whether the model degrades gracefully at regime boundaries, whether uncertainty estimates are calibrated, and whether the feature set is stable enough to remain informative over rolling one-year windows.
The Feature Set Problem Comes First
Before reaching the model architecture question, the feature set question dominated a large portion of our early work. Vol forecasting is seductive for ML practitioners because the raw input space is large: realized vol at multiple windows, implied vol across strike and expiry, term structure slope, skew metrics, cross-asset signals, macro data releases, news sentiment scores. The temptation is to feed all of it into a flexible model and let it find the structure.
The problem is that financial time series are not stationary. Features that were predictively relevant during one market regime become noise or confounders during another. A model trained on a feature set that includes variables correlated with the 2020 vol spike will learn regime-specific patterns that have no generality. On a rolling out-of-sample window, this shows up as overfit: in-sample error is low, out-of-sample error is systematically higher and not explained by model complexity alone.
Our feature engineering process applies a stability filter before anything is passed to any model. We estimate the predictive information contribution of each candidate feature on a rolling walk-forward basis, not on the full sample, and only retain features whose contribution is consistent across rolling windows. This eliminates a significant fraction of candidates that look promising in sample but are regime-specific leakage. The remaining feature set is smaller and less spectacular than the full candidate pool, but it produces more honest out-of-sample evaluations.
GARCH-Class Models as the Baseline
The appropriate baseline for any ML vol forecasting model is not a naive historical average. It is a well-specified GARCH-class model, which is a strong baseline in financial vol forecasting. We use GJR-GARCH with asymmetric volatility response (negative returns drive vol higher than same-magnitude positive returns) as the primary econometric benchmark. This is a decades-old model class that remains competitive on out-of-sample data against many ML alternatives, particularly at 1 to 5-day horizons.
GARCH does fail in specific situations: structural breaks, rapid regime changes where past vol is a poor predictor of near-term vol, and multi-horizon surface forecasting where you need joint predictions across strike-expiry grids rather than a single variance number. Knowing where GARCH beats ML models in walk-forward tests is a useful calibration check. Any ML model that cannot consistently outperform GARCH-GJR on 5-day realized vol prediction on out-of-sample data has not earned its place in the pipeline.
Random Forest and Gradient Boosting: Reliable but Capacity-Limited
Tree-based ensemble models, specifically gradient boosted trees with appropriate regularization, performed better than deep learning architectures on our shorter-horizon vol prediction tasks (1 to 10 days forward). The reasons are not surprising in retrospect. Tree-based models have lower sample complexity than deep networks, tolerate heterogeneous feature scales without much preprocessing, and have better-characterized failure modes that are easier to detect through validation monitoring.
For the feature set we retained after the stability filter, gradient boosted trees on a rolling 500-day training window outperformed GJR-GARCH on near-term realized vol prediction in conditions that included calm regimes and moderate regime transitions. The improvement was material on transitions: GARCH anchors to recent vol history, while tree models could use cross-asset features to detect regime shifts slightly earlier.
The limitation of tree-based models in this context is the output space. They produce point forecasts, or at most quantile estimates, for a scalar target. Extending them to forecast an entire surface (multiple strikes and expiries simultaneously) requires architectural decisions around how to handle the joint output structure. We use a combination of single-target models run in parallel with a correlation correction layer applied post-prediction, which is functional but has overhead.
LSTM and Transformer Architectures: High Variance on Vol Data
We tested LSTM networks and a compact transformer architecture on the same walk-forward setup. The results were mixed in a specific way that is worth documenting.
LSTM models showed competitive performance on 1 to 3-day forward prediction tasks when trained on abundant data with careful regularization. On shorter histories (less than two years of training data for a given asset), they overfit consistently. The capacity for sequence modeling that makes LSTMs useful in NLP and some signal processing contexts does not translate cleanly to daily vol data, which has roughly 250 observations per year. On our walk-forward protocol, LSTMs required at least 36 months of training data to produce stable out-of-sample estimates, and even then their improvement over tree-based models was marginal on average and higher in variance.
Transformer architectures, applied to vol prediction with attention over the feature time series, showed interesting in-sample behavior that did not hold up out-of-sample. The attention mechanism can learn spurious correlations that are internally consistent within the training window but do not generalize. On a 60-day rolling out-of-sample evaluation, transformer models had more outlier prediction errors during regime transitions than tree-based models, the opposite of what we hoped to find.
We are not claiming that deep learning cannot work for vol prediction. We are saying that in our specific setup, with the data volumes available and the evaluation protocol we used, tree-based models were more reliable. That finding is specific to this problem configuration and should not be generalized carelessly.
Where ML Adds the Most Value: Cross-Asset Signal Integration
The clearest advantage we found for ML models over traditional econometric approaches was in multi-signal, cross-asset integration tasks. Predicting tomorrow's single-name equity realized vol from that stock's own vol history: GARCH is competitive. Predicting the vol surface shift in equity index options by integrating signals from rates vol, FX positioning, and macro surprise data simultaneously: ML models were substantially better at maintaining predictive information from the full signal set.
The reason is structural. GARCH-class models are single-equation, single-asset models by design. Extending them to multi-equation systems (VAR-GARCH and related models) is possible but becomes unwieldy with more than three or four cross-asset inputs. ML models, particularly tree-based ensembles, handle large feature sets with heterogeneous information content more naturally.
This is where the ML layer in the Metafide pipeline earns its place: not as a replacement for parametric vol models, but as the integration layer that combines cross-asset signals with the parametric vol estimates to produce a richer input for surface construction. The output of the ML layer is not a final surface; it is an adjusted set of inputs to the parametric surface fitting step.
Walk-Forward Discipline: The Non-Negotiable
Every comparison we ran used strict walk-forward out-of-sample evaluation with a gap between training and test windows. No lookahead. No using any information in the test period to select hyperparameters or feature sets. This discipline significantly reduced the apparent advantage of every model relative to full-sample cross-validation results.
The gap matters: even one day of lookahead in a vol forecasting context can produce artificially strong results because vol is autocorrelated. A model that inadvertently uses future information will appear to predict vol well because vol today is correlated with vol yesterday, and if the model can see tomorrow, it anchors to tomorrow's value.
The walk-forward results are less impressive than full-sample results. That is intentional. A vol forecasting tool that overstates its predictive performance in backtesting and then underdelivers in live operation is worse than one that honestly characterizes its limitations. The Metafide platform communicates what the model can and cannot predict, and that communication is grounded in the walk-forward evaluation, not the in-sample fit.
This article is research analysis only and does not constitute investment advice. Metafide does not manage money or execute trades.