Beyond the Smile: A Hybrid Convolutional VAE for Crypto Volatility Surfaces

2026-06-15 · Source: Machine Learning · Field: Finance & Economics — FinTech & Digital Financial Services, Capital Markets & Investment Management, Artificial Intelligence & Machine Learning · Depth: Expert, quick

Summary

A novel hybrid convolutional variational autoencoder (VAE) is introduced for predicting and completing cryptocurrency implied-volatility surfaces. This model, which integrates a quadratic smile re-fit via a deterministic per-tenor routing rule, was trained on 6,034 hourly Binance Options surfaces for BTC and ETH from May-October 2023, using a \$6 \times 7$ tenor-delta grid. The VAE achieves a hidden-cell surface-completion RMSE of 0.94-1.56 vol-points across 10-50% mask rates. The hybrid predictor significantly outperforms a standalone smile re-fit, achieving 0.83 vol points at 50% masking compared to 7.00, an eightfold reduction. Crucially, it maintains 1.5-1.9 vol points error under structurally-correlated hole patterns where the smile re-fit incurs 9.6-13.1 vol points. Joint training on BTC and ETH improved performance by 9-27%, suggesting a shared volatility manifold. The hybrid model also ensures calendar- and butterfly-arbitrage-free surfaces and identifies market events like the late-October ETF-anticipation rally and the August 17, 2023 flash crash through reconstruction error.

Key takeaway

For quantitative analysts or machine learning engineers building volatility models for cryptocurrency options, you should consider integrating hybrid generative models like the proposed VAE. This approach significantly improves surface completion accuracy, especially in sparse data scenarios, and ensures arbitrage-free surfaces. Your models will gain robustness against structural data gaps and provide an unsupervised signal for market events, enhancing risk management and trading strategy development.

Key insights

A hybrid VAE effectively completes crypto volatility surfaces, outperforming traditional methods and identifying market anomalies.

Principles

• Generative models excel in sparse, structurally-correlated data regimes.
• Joint training across related assets can reveal shared underlying manifolds.
• Arbitrage-free constraints enhance model robustness and realism.

Method

A convolutional VAE is combined with a quadratic smile re-fit using a deterministic per-tenor routing rule. This hybrid model is trained on historical options surfaces to predict and complete implied volatility.

In practice

• Implement hybrid VAEs for robust volatility surface completion.
• Use reconstruction error as an unsupervised market anomaly detector.
• Explore joint training for related financial instruments.

Topics

• Convolutional VAE
• Implied Volatility Surfaces
• Cryptocurrency Options
• Financial Machine Learning
• Arbitrage-Free Modeling
• Market Anomaly Detection

Best for: AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• Yield Curves Dynamics Using Variational Autoencoders Under No-arbitrage — A physics-informed generative model forecasts yield curves with high accuracy and no-arbitrage compliance.
• HKUDS / Vibe-Trading — Vibe-Trading empowers users to transform natural-language finance queries into executable trading analysis and bounded autonomous trading.
• VAEs: Explained Once and for All — VAEs are generative models designed to approximate data's true probability distribution, with the autoencoder part being a practical means to an end.
• Volterra Generative Models — Volterra generative models use path-dependent noise for score-based generation, addressing non-Markovian dynamics with Markovian lifts.
• Forecasting the U.S. Treasury Yield Curve: A Distributionally Robust Machine Learning Approach for Interest Rate Risk Management — Integrating machine learning with distributionally robust optimization yields stable, tail-aware financial forecasts under market uncertainty.
• Hölder++: Improving the Quality-Coherence Trade-off in Multimodal VAEs — Hölder++ enhances multimodal VAEs by resolving the quality-coherence trade-off through exact Hölder pooling and disentangled representations.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning.