Keith So

Designed and deployed the desk's crude data platform ingesting 41 datasets into a versioned DuckDB database served by a 32-page Streamlit application. Developed the landed-cost arbitrage engine and refinery netback layer covering 39 physical crude routes with 78-84% entry signal hit rate. Delivered a live Brent-Dubai spread forecast using machine learning with 0.44 walk-forward IC. Implemented physical market and positioning monitors including Platts North Sea window, forward curves, and CTA replica tracking.

Designed and priced levered proceeds asset swaps with a Streamlit dashboard for UK gilts/linkers analysis. Delivered strategic asset allocation for multi-asset portfolios using Ortec Finance with multi-objective optimisation. Developed two dashboards: a real-time market analysis app in Shiny for Python and a Streamlit peer analysis app on SFCR data. Built a Solvency II Market Risk model in Python reducing computation time by 80%.

Identified and monetised short- and long-term derivatives trading opportunities across discretionary, systematic and option books while managing risk and delta/gamma hedging. Generated +260k PnL on 50k VaR in 2023 and +600k on 150k VaR in H1 2024. Researched, developed, and executed low-frequency quantamental models across 13 commodities using Bloomberg, SMM, Fastmarkets, and Oceanbolt data. Developed ML models achieving backtest portfolio Sharpe ratios of 1.5-2.

Developed volume and price-direction ML models to beat VWAP pricing in half-hourly energy markets achieving over 80% accuracy with Ridge regression. Built a Python backtest engine that supported a new product launch.