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

Battle Count: 74

Directly addresses core operational needs of options market-making and risk-management desks: surface completion under partial/missing quotes, no-arbitrage compliance, feed-disruption handling (entire tenor withdrawal), and cross-asset volatility modeling. The hybrid predictor is deployable at no additional inference cost beyond the standard smile re-fit. The 8x error reduction at 50% masking and order-of-magnitude improvement under structured holes (row withdrawal) are practically significant for production systems. The unsupervised anomaly signal provides a free monitoring tool for regime transitions. The work is explicitly framed around practitioner operational requirements.

The ConvVAE architecture is straightforward: three 3x3 convolutional layers with GELU activations, 16-dimensional latent space, standard reparameterization trick. Training completes in 10-50 seconds on a consumer GPU. The hybrid routing rule is a simple deterministic threshold (>=3 observed cells per tenor row). The smile re-fit baseline is a standard OLS quadratic fit. The main complexity lies in the data pipeline: parsing Binance EOH archives, put-call parity forward inversion, Newton iteration for IV recovery, quadratic smile fitting, delta inversion via fixed-point iteration, and constructing the 6x7 tenor-delta grid. All code is released and the complete ablation grid runs in under 5 minutes.

All training and evaluation infrastructure is released under MIT License on GitHub. Processed 6x7 gridded volatility surfaces, per-run configurations, checkpoints, and metric files are deposited on Zenodo (DOI: 10.5281/zenodo.20693546). The sole data source (Binance Options EOH archive) is publicly available. Complete ablation grid trains and evaluates in under five minutes of GPU time on commodity hardware. Fixed-seed masks ensure paired comparisons across methods. Time-ordered 70/15/15 splits are explicitly defined.