Gl4d3/shamba-signal

# 🌱 Shamba Signal **An honest machine-learning study of Kenya's county-level maize yields.** Shamba Signal asks a simple question: > Can a small set of annual ERA5 weather features improve maize-yield prediction enough to beat transparent historical baselines? The answer was **no—not enough**. Instead of hiding that result or searching endlessly for a better-looking model, the project publishes the no-go and makes the evidence explorable. ## 🎯 The result The study reconciled a private panel of **564 county-year rows** across all **47 Kenyan counties** for 2012–2023. Of those, 563 yield labels were usable. The temporal split was fixed before evaluation: - **2012–2021:** training - **2022:** model selection - **2023:** one-time final test; official labels are provisional | Model on provisional 2023 | MAE t/ha ↓ | RMSE t/ha ↓ | | --- | ---: | ---: | | **County historical mean** | **0.2998** | **0.3982** | | Weather Ridge | 0.3370 | 0.4537 | | Temporal Ridge | 0.3615 | 0.4783 | | Previous year | 0.4651 | 0.6057 | Weather Ridge improved on Temporal Ridge, but remained **0.0372 t/ha behind** the county historical mean on MAE. The scientifically correct decision was therefore **no-go** for this feature set. ## 🧠 What the ML pipeline does ```mermaid flowchart LR A[Official county maize data] --> B[Reconciled 47-county panel] W[Open-Meteo ERA5] --> C[4 annual weather features] B --> D[Leakage-safe temporal split] C --> D D --> E[Previous year] D --> F[County historical mean] D --> G[Temporal Ridge] D --> H[Weather Ridge] E --> I[Provisional 2023 test] F --> I G --> I H --> I I --> J[FastAPI evidence dashboard] ``` The weather model uses: - annual precipitation total; - wet-day count, where daily precipitation is greater than 1 mm; - annual mean 2 m temperature; and - annual maximum 2 m temperature. Same-year production and harvested area are excluded because yield is derived from them and including them would leak the answer. ## 🖥️ What the dashboard can …