cnemri/world-cup-2026-predictor

This project is a machine learning system that forecasts the results of the 2026 FIFA World Cup, built by training on over 150 years of international football matches going back to 1872. Instead of just guessing a winner, it produces calibrated probabilities for how likely each team is to win, based on a combination of models. The system starts with an Elo rating engine, a well known method for ranking teams based on match results, adjusted here so that World Cup finals count for more than friendly matches. On top of that sit a gradient boosted classifier that predicts match outcomes and a pair of models that predict the expected number of goals each side scores. There is also an optional model that blends in a newer type of AI model designed for tabular data. The README is careful to explain that every feature used to predict a match is only built from information available before that match kicked off, so the system cannot accidentally learn from future results it should not know about yet. The project takes evaluation seriously: it retrains itself repeatedly on expanding historical windows and tests each version only on matches it has not seen, which is described as the honest way to test a forecaster rather than testing on data the model already learned from. It reports several statistical accuracy measures rather than a single number, and compares its own performance against simpler baselines like an Elo ratings only approach. Once trained, the system runs a Monte Carlo simulation, replaying the entire tournament bracket fifty thousand times with randomly drawn scorelines, to estimate each team's overall chance of winning the whole tournament. Everything runs from a command line tool with separate steps for downloading data, building features, training, backtesting, simulating, and generating a final report, and the project includes its own test suite.