About

This notebook analyzes 8,494 historical settled weather markets from Kalshi's KXHIGHNY series (NYC daily high temperature) to quantify and correct systematic miscalibration in market prices.
What We Found:
Raw market probabilities have an Expected Calibration Error (ECE) of 0.01624
Using Isotonic Regression, we recalibrated them to ECE 0.00109
14.8x improvement in calibration accuracy
The Analysis:
Fetch all settled KXHIGHNY markets via Kalshi's public API (no auth required)
Extract market prices and outcomes across 8,494 trades
Bin predicted probabilities into 10 deciles
Compare predicted vs actual outcome rates — find where markets systematically over/under-estimate
Train Isotonic Regression, Platt Scaling, and Beta Calibration models
Isotonic Regression wins with best calibration performance
Visualize the calibration curve against perfect diagonal
Why It Matters:
Weather markets are a proxy for real uncertainty. If Kalshi is systematically wrong about temperature, traders pricing weather derivatives are wrong, insurance products are mispriced, and hedging strategies are based on false probabilities. This tool recalibrates raw market prices to truth.
The Workflow:
Data ingestion → exploratory analysis → feature engineering → model training → cross-validation → visualization → insight extraction. All in one reproducible notebook. No external files, no pickle dependencies, full transparency.
For Judges:
Click into any block and re-run it. The data fetches live. The models train on demand. The visualizations regenerate. This is science you can verify.