A machine-learning project to forecast monthly rainfall for Mumbai, India, using 121 years of historical data (1901 – 2021). Built for Spinnaker Analytics as part of the Data Science & AI track.

Mumbai receives ~90% of its annual rainfall during the four-month southwest monsoon (June – September). Anticipating monthly rainfall — even a few months ahead — helps the city's water utility plan reservoir releases, schedule infrastructure maintenance, and trigger demand-management protocols before crisis hits.

This project benchmarks five forecasting approaches and ships a working 12-month forward forecast.

XGBoost wins by a wide margin — a ~69% RMSE reduction over the SARIMA baseline. The strongest single feature is lag_12 (rainfall in the same calendar month a year ago), which carries ~48% of the model's predictive power.

Rain Forecasting/
├── rain_forecasting.ipynb              # Full Jupyter notebook (EDA + modeling)
├── rain_forecasting_report.docx        # Final report (Word document)
├── rain_forecasting_presentation.pptx  # Final presentation (PowerPoint)
├── mumbai-monthly-rains.csv            # Source data — Mumbai monthly rainfall
├── rainfall_forecast_next_12_months.csv# 12-month forward forecast (XGBoost)
├── feature_importance.csv              # Ranked feature importances
├── xgboost_rainfall_model.pkl          # Trained XGBoost model
├── sarima_model.pkl                    # Trained SARIMA model
├── forecast_plot.png                   # Forecast visualisation
├── link.txt                            # Source URL for the dataset
└── README.md                           # This file

• Source: OpenCity India — data.opencity.in/dataset/mumbai-rainfall-data
• Period: 1901 – 2021 (121 years)
• Frequency: Monthly
• Records: 1,452 monthly observations
• Target: Rainfall (mm)
• Annual mean: ~2,150 mm

1. Load the wide-format CSV.
2. Drop the Total column; melt to long format (Date, Month, Rainfall).
3. Exploratory analysis — time-series plot, monthly seasonality boxplots.
4. Stationarity check via Augmented Dickey-Fuller (raw + first-differenced).
5. Feature engineering — lag_1 … lag_12, Month, Rainfall_diff.
6. Time-based train-test split — train 1901 – 2010, test 2011 – 2021.
7. Train SARIMA, Random Forest, XGBoost, Prophet, and an LSTM.
8. Evaluate on the held-out test window using MAE and RMSE.
9. Generate a 12-month iterative forward forecast with the best model.
10. Persist the trained models and forecast outputs to disk.

pip install pandas numpy matplotlib seaborn statsmodels scikit-learn xgboost prophet tensorflow

1. Open rain_forecasting.ipynb in Jupyter / VS Code / Colab.
2. Run cells top to bottom.
3. The notebook will regenerate feature_importance.csv, rainfall_forecast_next_12_months.csv, and the .pkl model files.

import pickle
import pandas as pd

with open("xgboost_rainfall_model.pkl", "rb") as f:
    model = pickle.load(f)

# Build the same feature row used during training:
# columns = ['Month', 'Rainfall_diff', 'lag_1', 'lag_2', ..., 'lag_12']
# then call model.predict(X)

• Climate non-stationarity — the model can't extrapolate beyond regimes seen in the 1901-2021 training data.
• Iterative forecast drift — long-horizon iterative prediction can shift the seasonal peak slightly.
• Univariate features — only past rainfall is used. Adding ENSO / IOD indices and SST anomalies would likely improve skill on anomalous years.
• Monthly resolution — flood-warning use cases need a daily model.
• Single train-test split — walk-forward cross-validation would give a more robust estimate.

• Use the 12-month forecast monthly to drive reservoir-release schedules across Mumbai's seven supply lakes.
• Auto-trigger demand-management protocols when forecast Jun-Sep volume falls below the 25th historical percentile.
• Concentrate pipeline and treatment-plant maintenance in months with forecast rainfall < 50 mm.
• Productionise the XGBoost model behind a forecasting API with monthly retraining and drift monitoring.
• Build a Power BI / Tableau dashboard showing forecast vs reservoir levels for monthly review meetings.

Daniyal Khan — daniyal.khan@growhut.in

Submitted to Spinnaker Analytics — April 2026.

• Per the project brief, the source dataset is not to be uploaded publicly (GitHub / Kaggle / etc.).
• All deliverables (notebook, report, presentation) should be zipped together for final submission.