Kaymata/market-price-predictor

Forecasts tomato and onion prices by state in Nigeria using WFP market data. 3MTT NextGen capstone AI-07. cat > README.md <<'MDEOF' # Market Price Predictor - Tomato & Onion **3MTT NextGen | Brief AI-07 | AI & Machine Learning cohort** Fellow: Kabiru Usman Bashir | Omni Digital Media Institute, FCT A forecasting tool for retail tomato and onion prices in Nigeria, built to help traders, food vendors and households answer one recurring question: buy now, or wait? --- ## What it does Pick a state and a commodity, and the app shows the recorded price history, a forecast one to six months ahead, and - critically - the range that forecast is likely to fall within. It also reports how volatile that market is, how prices compare to a year ago, and whether a seasonal pattern exists at all. ## The headline finding I set out to model seasonality: tomato prices are supposed to spike in the rainy lean season and fall at harvest. **In this data, they do not.** Measured on the cleaned series, the calendar month explains only **2% of month-to-month price movement** in Borno and Yobe. A seasonal-naive forecast (assume this month repeats last year's move) loses to simple persistence by **44%**. These are conflict-affected markets: prices are driven by road closures, insecurity and fuel costs, not the agricultural calendar. So the tool tests three candidate forecasts on held-out data - a trained gradient-boosting model, a drift term, and naive persistence - and uses whichever wins. Here, persistence wins, and the tool says so rather than manufacturing a trend. The forecast is presented with an empirical uncertainty band, because in a market that moves +/-16% a month, the range is the honest answer and a single number is not. ## Results Tested on 21 observations from January to June 2026, never seen during training (3,340 training rows). | Forecast | MAE (N per 0.5 kg) | MAPE | |---|---|---| | Naive persistence baseline | 36.00 | 12.6% | | This system | 36.00 | 12.6% | | Learned weights | model 0.00, drift 0.00 | | The model weight of zero is the result, not a bug: the system …