SurfaceEdge - Options Pricing with Novel CNN

SurfaceEdge is a novel multimodal deep learning pipeline for predicting next-day options contract price changes. Existing approaches treat each contract in isolation using scalar features, or apply LSTMs for time-series dynamics. Neither approach leverages the full spatial structure of the options surface. SurfaceEdge addresses this by encoding each day's options chain as an 60x30 RGB image (implied volatility, open interest, volume) and passing it through a CNN backbone, fused with contract-level scalar inputs.

Built on a dataset of 200M+ labeled contracts spanning 17 years across 104 U.S. equity tickers, we designed and compared three architectures: a baseline multimodal CNN, a deep residual variant, and a causal self-attention transformer that models price history as a sequence prediction problem, similar to next-token prediction in LLMs. The transformer finished as our top model, achieving a 7.9% improvement in MAE over the naive baseline.

This project extended my prior computer vision research with CU as an RA, applying CNN-based image recognition to a quantitative finance problem and combining it with the market intuition developed through my work on the CU Quants trading fund.