Prediction-market research

This started as a calibration exercise: could I measure which information sources actually lead consensus, and price outcomes against that instead of against the news cycle?

The core tool is a Python NLP pipeline that scores sources by their measured historical reporting lead-time versus wire-service confirmations, with strict source-tier qualification rules for what text is even allowed into the model. During fast-moving events, the edge is not secret information; it is knowing which public source tends to be right, and how many hours early.

Two pieces of method mattered more than any single trade. First, operationalizing fuzzy resolution criteria: a contract like "out of power by [date]" becomes explicit decision rules — death, incapacitation, or formal replacement — before any money is at risk. Second, modeling state actors by treating publicly announced positions as negotiating ceilings, which is how I priced the continued closure of the Strait of Hormuz from May through August and the February 2026 strike on Iran, the event that accounts for about $10K of the total.

Position sizing is by expected value: laddered entries reaching 25–50% at peak conviction, with explicit drawdown limits on intrinsically noisy negotiated-outcome markets.

The honest part: I wrote a formal post-mortem on the negotiated-ceasefire event class, where this method failed. Ceasefire announcements are designed to be believed before they are true, and lead-time on the announcement is not lead-time on the outcome. The lesson I took from it: on markets whose resolution depends on a signature rather than a fact on the ground, calibrated estimation has to include the announcement's own incentive structure, not just the reporting speed of the sources covering it.

Roughly 400 hours of methodology work so far, across ~15 linked geopolitical questions. The account grew from about $1K in total deposits to $17,500 in about four and a half months. The markets are still open, and so is the question of how much of that was the method.