This project forecasts weekly sales for 45 Walmart stores by combining three years of historical sales with seven U.S. economic indicators from the Federal Reserve's FRED database, such as consumer sentiment, inflation and interest rates.

I built it as a complete, reproducible pipeline that downloads and merges the data sources automatically, engineers seasonal and trend features, and trains and compares five machine learning models against a simple baseline.

I designed it the way production forecasting systems need to work. Models are always trained on past weeks and tested on later weeks, and every feature only uses information that would have been available at the time, so the results reflect real forecasting performance rather than hindsight.

The best model predicts weekly store sales within about 5% on average, compared with 67% for the baseline. The codebase follows software engineering practice, with modular components, a shared model interface that makes adding new models simple, logging, and an automated test suite. Next, I'm adding time-aware hyperparameter tuning and model interpretation to explain which economic factors drive store sales.