Logo Lanfrica

vkakorsu/farmsmart-ghana

Domaine:

agriculture

Type de record:

software
Créateur:
vka
Hôte:
Crop price intelligence and sell-timing advisory for Ghanaian smallholder farmers - Ghana AI Innovation Challenge 2026 # FarmSmart Ghana Crop **price intelligence and sell-timing advisory** for Ghanaian smallholder farmers, built on real Ghanaian market data for the Ghana AI Innovation Challenge 2026. **Live demo:** farmsmart-ghana.streamlit.a… Smallholders lose income two ways: they sell at the **wrong time** (dumping at harvest when prices bottom) and in the **wrong market** (selling locally when a nearby market pays far more). FarmSmart answers two questions in plain language (and, in the full build, over USSD/SMS/WhatsApp in local languages): - **When should I sell?** Seasonal forecasting of storable-staple prices. - **Where should I sell?** Cross-market price-gap (arbitrage) detection. ## Primary Ghanaian dataset World Food Programme **Ghana Food Prices**, published on the Humanitarian Data Exchange (HDX): monthly retail/wholesale prices across Ghanaian markets, in GHS. The pipeline downloads it automatically via the HDX CKAN API. See docs/DATA.md for full source disclosure and complementary sources (MoFA SRID, GSS, FSNMS). ## Results so far (real backtest, fully reproducible) Data after cleaning: **33,369** per-kg price records, **16 storable staples**, **19 markets**, **2007-2023**. Backtest = strict time split, test on the last 18 months. Prices normalised to **GHS per kg**; perishables excluded (sell-timing only helps crops a farmer can store). Forecast accuracy by horizon (median APE, lower is better): | Horizon | Naive (random walk) | FarmSmart (LightGBM) | | --- | --- | --- | | 1 month | **6.3%** | 12.3% | | 3 months | **17.1%** | 21.3% | | 6 months | 28.8% | **27.5%** | Honest reading: at short horizons the random walk is a strong baseline (prices are persistent). At the **6-month horizon - the one that matters for storage and sell-timing decisions - FarmSmart beats both naive baselines on every error metric** (MdAPE, MAPE, MAE, RMSE) and beats naive on 53% of individual forecasts. Full metrics and the insight tables are in results/ (`metrics.csv …

Licenses