Priors
A 24/7 advisory engine for Kalshi speaker-mention markets — hierarchical Bayesian pricing on live transcripts, published with its own negative results as the headline.
The problem
"Will the speaker say this phrase" markets price on vibes. Can live transcripts plus disciplined priors find real mispricings — and can you prove it either way?
The approach
Five concurrent service loops run the system: a market watcher polls Kalshi on a 30-second production cadence, and ~1,800 phrases are matched against live transcripts with negation- and attribution-aware boundary matching. Each (speaker, phrase) pair gets a hierarchical Beta-Binomial posterior — thin data shrinks toward the speaker’s base rate, outcomes decay on a 60-day half-life — and the engine flags a bet only when the 90% credible interval excludes the market price. Advisory-only is architectural: the codebase contains zero order-placement calls.
Highlights
- Hierarchical Bayesian pricing with credible-interval betting gates and Kelly sizing
- An append-only decisions log where every model change must cite outcome-level evidence
- A signal learner that drove its own losing signals’ weights to zero in production
Results
- Ran live for months: 1,115 resolved bets, and the README leads with the loss — $16.40
- Measured its own Brier score as worse than the market, then rebuilt the model on that evidence
- ~40,000 lines of Python, 264 tests passing in about ten seconds, public under MIT
The discipline is the product: instrument everything, publish the result you got — then build the next model on evidence.