Sports market machine learning

Sports Prediction Mispricing

Machine learning and time-series models analyzing how NBA prediction markets respond to major in-game win-probability shocks.

View GitHub Back to portfolio

Background

Do prediction markets react rationally to major events during NBA games?

Prediction-market prices reflect traders’ expectations about real-world outcomes, but emotionally charged moments may cause markets to respond inefficiently. In NBA games, high-impact plays can produce sudden changes in win probability while simultaneously driving trading activity on prediction markets.

We combined Kalshi prediction-market data with ESPN game and win-probability data to study how market prices adjust after influential plays, whether those adjustments reveal overreaction and mispricing, and whether short-term market responses can be predicted from recent trading behavior.

Project Goals

Identify High-Impact NBA Events

Use changes in ESPN win probability to identify influential plays that provide natural event shocks for studying prediction-market behavior.

Measure Market Reactions

Align each win-probability shock with Kalshi market candles and quantify how prediction-market prices respond immediately after the event.

Detect Potential Mispricing

Compare changes in win probability with corresponding market movements to distinguish appropriate reactions from potential overreactions.

Predict Short-Term Market Behavior

Compare traditional machine-learning baselines with LSTM and Transformer sequence models for predicting market direction and price one minute after a shock.

Project Workflow

01 Identify Win-Probability Shocks

Use ESPN play-by-play data to identify NBA events associated with changes of at least 5% in game win probability.

02 Align Market and Game Data

Match ESPN events to Kalshi market candles and filter sequences for sufficient temporal coverage and data quality.

03 Construct Event Sequences

Build 20-minute windows around each shock so models can observe how market prices behave before and immediately after influential plays.

04 Classify Market Direction

Train Logistic Regression, Histogram Gradient Boosting, LSTM, and Transformer classifiers to predict the direction of the market one minute after an event.

05 Forecast Market Price

Use linear, gradient-boosting, LSTM, and Transformer regression models to predict the Kalshi contract price one minute after a win-probability shock.

Figures

Selected outputs from the analysis.

Line chart comparing model predictions with true post-shock market price for a held-out event. — **Held-out event sequence**Sequence models track the immediate post-shock price path more closely than tabular baselines.

Scatter plot comparing ESPN win-probability shocks with Kalshi price changes around NBA events. — **Win probability shock vs. market move**Most market moves do not scale cleanly with shock size, and true one-minute overreactions are rare.

Classification figure distinguishing underreaction, efficient response, and overreaction windows. — **Mispricing classification**The labeling frame separates slow updates from genuine overreactions rather than treating every price gap as the same kind of error.

Key Result

Sequence models accurately predict short-term Kalshi responses to influential NBA events.

The strongest classification model, the LSTM, achieved 82.2% balanced accuracy when predicting one-minute market direction. For price forecasting, the LSTM achieved an R² of 0.994, outperforming both conventional machine-learning baselines and the Transformer.

Together, the results show that incorporating the temporal structure of prediction-market prices around NBA win-probability shocks substantially improves short-horizon market prediction.

Associated Presentation

Predicting NBA Mispricings: CS 1090B Final Presentation

Final project presentation explaining the Kalshi and ESPN data join, shock-window labeling, sequence-model setup, and overreaction findings.