Sports market machine learning

Sports Prediction Mispricing

Kalshi NBA prediction-market analysis joined to ESPN win-probability shocks to test whether markets overreact or underreact in game.

View GitHub Back to portfolio

Background

NBA prediction markets can move quickly after major in-game win-probability swings, but it is unclear whether those moves systematically overreact to new information or update too slowly.

This project studies whether NBA prediction markets systematically misprice sudden in-game information. It joins Kalshi KXNBAGAME market candles with ESPN play-by-play win-probability shocks from the 2026 NBA playoffs, then asks whether one-minute price moves overreact, underreact, or update efficiently after major plays.

Read this page as a market microstructure and sequence-modeling project: the analysis defines event windows around win-probability shocks, labels market response behavior, and compares tabular baselines with models that can read the local 20-minute price path.

Project Workflow

Define a market response window

Use ESPN play-by-play win-probability jumps as event anchors, then inspect Kalshi one-minute prices around each shock.

Build leakage-safe sequences

Group train/test splits by game and represent each shock-market pair as a fixed-length OHLCV sequence so models cannot memorize games across splits.

Separate overreaction from slow updating

Compare the market price move against the win-probability shock to distinguish rare large overreactions from the more common underreaction pattern.

Compare tabular and sequence models

Benchmark logistic and linear baselines against LSTM and Transformer models that can use the full local price path.

D3 Views

Native charts make the market-response results easier to scan.

The sharper finding is that true one-minute overreactions are rare, while models that can read the local OHLCV sequence do a much better job predicting immediate direction.

Label Rarity

Only a tiny share of sequences meet the strict overreaction rule.

The README reports 21 true one-minute overreactions among 12,422 event-market sequences.

Share of all event-market sequences.

Model Comparison

Sequence models outperform tabular baselines on direction.

Balanced accuracy controls for the up, flat, and down classes.

Held-out balanced accuracy by direction model.

Event Window

A 20-minute price path anchors each shock.

The example shows ten minutes before the play, the event minute, and the immediate post-shock response.

Kalshi yes-price around a win-probability shock.

Figures

Selected outputs from the analysis.

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.

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.

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.

Associated Report

CS 1090B Sports Mispricing Final Report

Final report covering the project motivation, Kalshi and ESPN data construction, model comparisons, limitations, and interpretation of market underreaction versus overreaction.