Financial Time Series Modeling — Algorithmic Trading Strategy Discovery
Cerebellum Capital was an algorithmic trading hedge fund based in San Francisco that discovered trading strategies using a mix of genetic algorithms, machine learning, and convex optimization. Any candidate strategy was backtested over long stretches of historical data to confirm it held up across very different market regimes rather than fitting one lucky period. The fund traded daily across a universe of 2,000+ stocks. Infocusp worked with them on modeling and infrastructure, involved end to end from strategy ideation through data curation to deploying the models and frameworks. This engagement ran from 2014 to 2019.
Strategy discovery: genetic programming + convex optimization
As an individual contributor on the team, I worked across the strategy-discovery stack. At the core was a genetic programming framework for searching the strategy space, fed with 200+ technical and fundamental features and a library of 100+ operations it could compose into candidate strategies, built to evaluate millions of those candidates in parallel. Paired with it was a convex optimization framework that maximized profitability while holding portfolio risk below defined thresholds. A large part of the value was in making that optimizer realistic: we translated the fund's real-world trading rules into convex constraints and encoded them directly into the formulation, so the allocations it produced were not just theoretically optimal but actually executable under those constraints. This improved their profitability by a margin.
Modeling & infrastructure
On the modeling side, we applied a range of time-series techniques — ARIMA models, RNNs and LSTMs, and transformers — to forecast future trends as inputs to the strategies. Supporting all of this was the infrastructure layer: cloud automation that scraped datasets, generated strategies, ran backtests, and surfaced final trade suggestions on a recurring schedule.
To keep strategy development fast and correct, we also built a library of vectorized, time-safe functions covering the operations traders reach for most often — with time-safe being the critical property, since it prevented lookahead bias from leaking future information into a backtest and quietly inflating results.