KM Kanav Mehta Quantitative AI Research

Prediction Markets · NLP

Pricing "will they say the word" like it deserves a model.

Kalshi lists binary and count markets on whether a word gets said during a speech or earnings call, and how many times. This is a fair-value engine for those contracts: a hurdle model, live updates, and walk-forward evidence against the base rate.

Back to Projects GitHub
Kalshi mentions dashboard overview with market counts, type breakdown, and mentions distribution
Overview: 2,533 mention markets, type breakdown, and the mentions-per-market distribution from the parquet snapshots.
Kalshi mentions pricing tab with hurdle model fair values
Pricing: hurdle-model fair values for mention contracts, straight off the transcript corpus.

The Model

Two questions, one prior.

  • First the hurdle: is the word said at all? Reference-class base rates (shrunk through a Beta-Binomial ladder from speaker×format down to speaker, format, and global word rate) give P(occurs) without overfitting thin history.
  • Then repetition: conditional on being said, a truncated negative-binomial layer models how many times, calibrated so E[N|N>0] matches history.
  • The product prices binary "said" markets and count ladders (≥N times) from one coherent distribution instead of separate heuristics.

Live Pricing

Updates as the speech happens.

  • Mid-event, a Gamma-Poisson prior is anchored to the hurdle estimate and updates continuously from elapsed words and observed counts. Exact binary continuity, no ad-hoc blending.
  • Opt-in, interpretable adjusters: an LLM reading the transcript corpus, and topic-similarity features over it.

Evidence & Interface

Beat the base rate, then show it.

  • Walk-forward backtests compare the model against the shrunk base-rate baseline, the honest benchmark for a market most participants price by gut.
  • The full stack ships as a CLI, a FastAPI backend, and a Mantine dashboard, all reading parquet snapshots of the mention universe.
  • Transcripts are plain JSON records (speaker, event type, date, text), so the corpus stays portable and auditable.