← All work
Machine learning · Open sourceArchived
Weather Trend Forecasting
A gradient-boosted temperature-forecasting system over 114K records from 204 countries, with an interactive map and a FastAPI backtester.
1.21 °C
Test MAE
XGBoost on the held-out test set.
204 countries
Coverage
114K records after cleaning.
Problem
Global temperature forecasting is a breadth problem: many countries, noisy seasonality, and a need to inspect predictions rather than trust one headline metric.
Approach
Careful data work first (211 raw country entries resolved to 204), then an XGBoost model with engineered temporal features, served through a FastAPI backtester with an interactive map, so any prediction can be checked against what actually happened.
Outcome
1.21 °C mean absolute error on the test set, and a tool for probing predictions country by country.