Rainfall_Forecast_Mausam — All-India Rainfall Risk Model

Hackathon MVP. A SINGLE pooled model, trained across 316 Indian districts simultaneously, predicting the probability of insufficient rainfall over the next 7 days for any of them, plus an expected total-rainfall estimate (mm). Replaces the earlier one-model- per-district approach — pull this once, use it for any trained district.

This is not an operational meteorological forecast. It does not claim panchayat-level accuracy and has not been validated by any meteorological authority.

Why a shorter (2021-2025) date range than the raw data

Restricted to 2021-01-01 onward: rainfall-reporting completeness has a hard, dataset-wide step change exactly at that date (missingness ~87% in Dec 2020 to ~6.5% in Jan 2021, uniformly across every district — a data-generation-process artifact, not a real gradual station-network rollout). Training only on the reliably-reported era avoids depending on an unverified imputation assumption for the bulk of the record.

Why no zero-imputation for missing rainfall (unlike the single-district models)

impute_as_zero=False for this pooled model — rows with an incomplete 7-day target window or 30-day feature lookback are dropped, not imputed. The Thiruvananthapuram-specific zero-imputation justification (missing correlates with warmer/drier conditions) was checked across 143 districts and did NOT replicate nationally (66 districts same-direction, 77 opposite-direction) — so no blanket imputation assumption is applied at this scale.

Target definition

• rainfall_next_7_days: sum of observed rainfall (mm) over T+1..T+7.
• insufficient_rainfall_next_7_days: 1 if that sum falls below the 33.0th percentile of that (district, calendar month)'s rainfall distribution, computed from the training split only (IMD-style lower-tercile convention) — a district- and-season-specific threshold, never a single national cutoff.

Data splits (chronological, no shuffling)

• Train: 2021-01-01 to 2023-06-30
• Validation: 2023-07-01 to 2024-06-30
• Test: 2024-07-01 to 2025-02-10

Model

• Classifier: xgboost ({'learning_rate': 0.1, 'max_depth': 5, 'n_estimators': 200}), calibrated (sigmoid (Platt), fit on validation split, cv='prefit' (FrozenEstimator))
• Regressor: xgboost ({'learning_rate': 0.05, 'max_depth': 5, 'n_estimators': 200})
• Class imbalance handling: RandomForest: class_weight='balanced'. XGBoost: scale_pos_weight =4.190 (train neg/pos ratio). No resampling/SMOTE.

Test-set performance vs. climatology baseline (pooled, all districts)

Baseline predicts only the historical (district, month-of-year) base rate — no information about recent/current weather. This is the bar the model must clear.

Metric	Climatology baseline	xgboost
Accuracy	0.790	0.754
Precision	0.000	0.394
Recall	0.000	0.318
F1	0.000	0.352
ROC-AUC	0.692	0.754
PR-AUC	0.299	0.379
Brier score	0.141	0.152

Regression (rainfall_next_7_days, mm):

Metric	Climatology baseline	xgboost
MAE	23.40	23.65
RMSE	41.25	41.93

Per-district sample (sanity check — the model isn't only good for high-row-count districts)

District	Test rows	ROC-AUC	Regression MAE
Thiruvananthapuram	211	0.494	28.45
Mumbai Suburban	210	0.895	38.88
New Delhi	210	0.678	26.77
Kolkata	209	0.838	24.21
Jaisalmer	200	0.995	13.78
Bengaluru Urban	198	0.667	18.35

Full metrics, including every sampled district: see evaluation_results.json.

Inputs / feature schema

See feature_schema.json. All features use only information available up to the prediction day; the 7-day forward window is only ever used to build the target. Geographic features (latitude/longitude/elevation) have real variance across training rows here (unlike a single-district model), so this model can actually use location, not just season, to predict.

Files in this repo

• rainfall_risk_classifier.joblib, rainfall_amount_regressor.joblib
• preprocessor.joblib — (district, month)-keyed climatology and risk- threshold tables, plus median-imputation values, all fit on the training split only
• climatology_baseline.joblib — the baseline model shown above
• feature_schema.json, model_metadata.json, evaluation_results.json

Limitations

• Trained only on districts present in the source dataset; a district not in model_metadata.json's districts list has no learned climatology/ threshold and should not be queried.
• Two district names collide across states in the source data (Raipur: Chhattisgarh/Madhya Pradesh; Cuddalore: Tamil Nadu/Puducherry) — these are disambiguated as "Raipur (CT)"/"Raipur (MP)" etc.
• Effective training window is short (~2.5 years) because of the data- completeness cutover explained above — less historical depth than the single-district models, compensated for by far more districts (breadth) per split.
• Not validated against independent ground-truth rainfall records.