Hi all and thank you to Zindi and everyone. My solution to the challenge was built on a simple but important insight: traditional rainfall prediction is not just about meteorology—it’s about people, patterns, and place. By treating the problem as one of behavioural modelling, I built a solution that didn’t just predict rainfall—it understood the cultural logic behind those predictions.

Approach The dataset presented a severe class imbalance: 88% of all entries were “NORAIN”, with the remaining three categories making up just 12%. Rather than treating this imbalance as noise, I treated it as signal. It reflected real-world behaviour: farmers naturally make more “no rain” predictions. This shaped my modelling strategy from the outset.

I focused on features that captured who was predicting, where, and when:

• User behaviour: Some farmers were 10× more active than others.
• Geographic specificity: Cleaned and standardized over 20 Ghanaian community names.
• Temporal rhythms: Rainfall intensity peaked midweek, especially on Wednesdays.
• Indigenous indicators: Lightning correlated with medium rain; sun/heat often preceded heavy rain.

This approach allowed the model to learn from behavioural and spatial patterns embedded in traditional forecasting practices.

Model / Code I compared three tree-based ensemble classifiers:

• XGBoost: Best balance between accuracy and generalisation using scale_pos_weight.
• LightGBM: Fast and efficient, with native support for categorical features.
• CatBoost: Ideal for categorical-heavy datasets like community and user ID.

All models were evaluated using stratified cross-validation and macro/weighted F1 scores. XGBoost emerged as the most stable and generalisable.

Final training setup:

from xgboost import XGBClassifier

Encode target labels

y_full_numeric = label_encoder.transform(y) X_features = X

Final model configuration

final_model = XGBClassifier( n_estimators=200, max_depth=5, learning_rate=0.1, random_state=42, n_jobs=-1, eval_metric='mlogloss', use_label_encoder=False, scale_pos_weight=scale_weights )

Train on full dataset

final_model.fit(X_features, y_full_numeric)

Evaluation The model performed reliably in identifying non-rain events but struggled with light rain. A modest 2.6% gap between training and cross-validation F1 scores indicated strong generalisation and minimal overfitting.

Using SHAP and LIME, I analysed what the model had truly learned:

• Global insights: User history and location were dominant predictors.
• Feature effects: Time features (hour, day) and most communities increased rainfall predictions. User ID and forecast length tended to decrease them. Clouds had a positive influence, but geographic and temporal context dominated.

This confirmed the model wasn’t just forecasting rain—it was decoding the behavioural logic of Ghanaian farmers.

This was an initial model designed to establish a clean, interpretable baseline. I did not apply any model enhancements like hyperparameter tuning or model stacking.

Thanks again to Zindi and the community. Hope this helps.