Personal project • Solo developer • 2026 • Live
Rwanda Inflation Forecaster
A public dashboard of Rwanda's economy with 12-month inflation forecasts, tested against what actually happened.
The problem
Rwanda's economic data is public but scattered across the IMF, the World Bank, WFP and weather archives, and inflation forecasts rarely show how accurate they've been.
What I built
- Daily ingest of 47 monthly series: consumer prices, exchange and interest rates, world commodities, exports, market food prices and provincial rainfall
- Inflation, food price, market and rainfall dashboards, each series showing its source and freshness
- 12-month inflation forecast with an 80% range from Google's TimesFM model
- An AI analyst that writes a briefing from the site's own figures, rate-limited and cached
Engineering decisions
Every method is backtested
At each of the last 72 months, each method forecasts using only the data available then. A no-change baseline, seasonal ARIMA, TimesFM, TimesFM with driver data and an average are compared on average miss and range coverage.
Honest results on the page
Seasonal ARIMA beat the TimesFM foundation model at every horizon, and the 80% range proved too narrow beyond a few months. The site says so.
Base effects handled correctly
The models forecast the price index level and inflation is derived from it, so past price jumps drop out of the year-on-year figure when they will in the real data.
Stack
What I learned
A newer model isn't automatically a better one. Without a backtest, I'd have shipped the foundation model and never known the classical one was more accurate.