ARCHIVES

This paper walks through an end-to-end options analytics pipeline I built from scratch—one that starts with synthetic data generation and ends with live trading signals on Deribit. The core idea was pretty simple: train machine learning models on simulated options data, then see if they could find real opportunities in the wild. What surprised me was how badly the initial approach failed on certain trade types, and what it taught us about the gap between textbook finance theory and actual market behavior. The pipeline combines XGBoost, neural networks (MLP, CNN, LSTM), and a tiered value-ranking system to flag trades worth taking. On synthetic data, things looked great—about 84.5% accuracy calling ITM/OTM and a modest but positive ROI. But when we went live, the mean-reversion bias baked into our synthetic generator caused some ugly losses on put options. That failure forced a rethink: we added momentum features like RSI and volatility skew, retrained everything, and the results improved dramatically. The real payoff came when the revised model caught a massive divergence in ETH options during a market dip—the kind of opportunity the original model would have missed completely.