Hal Beresford · Forecasting systems
There's a class of question with a real answer we just don't know yet.
Will some measurable event happen, and by when? Prediction markets let people trade on those questions, and the prices act like a crowd's best guess at the odds.
For the past several months I've been building a system that competes in those markets — on Kalshi, a regulated exchange for event contracts. It gathers public data across a range of domains, turns it into calibrated probabilities, and, where its estimate disagrees with the market's price, places a trade. Then it settles up, scores how it did, and learns from the record.
It runs by itself: about thirty forecasting pipelines operating unattended on a single server, all day, alerting me only when something breaks. Building it touched everything — data engineering, statistical modelling, live financial execution, and the unglamorous reliability work that keeps an automated system from quietly going wrong.
What it involves
Gather
Public data from many sources, cleaned and made point-in-time honest so a backtest can't peek at the future.
Forecast
Probabilistic models, calibrated and scored against real outcomes rather than trusted on faith.
Execute
Live orders placed and reconciled automatically, built to fail safely when something goes wrong.
Open source
The forecasting itself stays private — but I've open-sourced the engineering underneath it: the exchange client that handles authentication, correct fee math, and crash-safe order submission. A clean, well-tested sample of how I build production systems.
View on GitHub →Get in touch
Looking for data science, product, or founding-team roles — anywhere the job is turning messy real-world data into decisions. Happy to walk you through the whole system.