Volatility in Prediction Markets: A Structural Approach

Battle Count: 67

Highly relevant for prediction-market trading, market making in binary contracts, and risk management of prediction-market portfolios. The structural volatility model provides interpretable forecasts of probability-move magnitudes, directly useful for adverse-selection risk in quoting, position sizing, and hedging. The finding that structural state variables (price, time to resolution, spread, volume) dominate generic GARCH dynamics by 34% is actionable for prediction-market traders. However, the paper focuses on prediction markets specifically rather than traditional asset markets, limiting direct applicability to equity/futures/forex trading. The GARCH+DR-AS hybrid framework could inform volatility modeling for any bounded-probability asset class.

The closed-form DR-AS model is relatively simple to implement: it requires only four state variables (price, time to resolution, spread, volume) and one fitted parameter K. The Wright-Fisher deadline-resolution term p(1-p)/tau is parameter-free. The GARCH+DR-AS hybrid adds standard GARCH(1,1) recursion with an additive structural baseline, requiring estimation of omega, alpha, beta, and c parameters via quasi-likelihood. The main complexity lies in the data pipeline: constructing hourly contract panels from Kalshi, handling contract boundaries, implementing expanding-window monthly estimation, and computing Winkler interval scores. The model does not require deep learning infrastructure or complex optimization.

The paper uses Kalshi contract data (Aug 2021 - Apr 2026) which is publicly accessible through the Kalshi platform. The structural model is fully specified with closed-form expressions. Estimation uses standard quasi-likelihood methods. However, no explicit GitHub repository or code release is mentioned. The hourly panel construction, filtering criteria (active-update definition, 48-hour minimum), and expanding-window monthly estimation design are described in detail. The Winkler interval score evaluation is standard. Reproducibility is moderate given the large dataset and specific filtering choices, but the methodology is transparent.