Climate riskTime series forecasting

Pakistan Rainfall Forecasting & Flood Early Warning

Can a century of rainfall data warn us about the next flood? SARIMA forecasting on 116 years of Pakistan monsoon data, evaluated against the 2022 disaster.

View codePortfolio write-up
SARIMA forecast with 2022 flood context
15%better RMSE than ARIMA
14.06 mmSARIMA RMSE
1,392monthly observations
5models compared

The problem

In 2022, monsoon floods in Pakistan displaced more than 30 million people and caused about $30 billion in damage, with roughly ten times normal rainfall.

This project asks whether classic time series algorithms, trained on more than a century of data, can recognize when rainfall is moving far outside its normal range.

Pakistan population, urbanization, agricultural land, and forest area over time
World Bank indicators: a growing, urbanizing population and shrinking forest cover raise exposure to floods.

Approach

Five models were compared across three datasets: monthly rainfall from 1901 to 2016, World Bank climate indicators, and NOAA's Oceanic Niño Index for El Niño and La Niña effects.

ARIMA(2,0,2) served as the non-seasonal baseline. SARIMA(1,0,2)(1,1,1,12) added the 12-month monsoon cycle. SARIMAX added ocean temperature data. Two naive baselines, the historical average and a repeat of last year, kept the models honest.

Results

ModelRMSE (mm)MAE (mm)
SARIMA(1,0,2)(1,1,1,12)14.0610.98
SARIMAX (+ El Niño data)13.7010.62
Historical average14.0911.01
ARIMA(2,0,2)16.6413.48
Last year repeat27.6318.38
Predictions versus actual rainfall for four models, with an error comparison chart
Each model’s predictions against actual rainfall, 2011–2018.

The warning signal

For July 2022 the model predicted about 57 mm of rain. The actual total was roughly ten times that. The forecast wasn't supposed to predict the flood. The point is that rainfall this far outside the confidence interval is the early warning.

SARIMA forecast with the 2022 flood season highlighted
SARIMA forecast with its confidence band, shown against the 2022 flood season.
El Niño sea surface temperatures versus monsoon rainfall
El Niño and La Niña conditions are correlated with monsoon rainfall (r = −0.258), but add little forecasting power.

Responsible AI check

Honest baselines. The naive historical average came within 0.03 mm RMSE of SARIMA. I report that openly: the model’s value is in its seasonal structure and uncertainty band, not a dramatic accuracy gain.

Extra data isn’t automatically better. El Niño data confirmed a real relationship with rainfall but barely improved the forecast, so the simpler model stays the headline.

Stated limits. National monthly totals miss river flow, local terrain, and daily intensity. The warning signal is a promising hypothesis, not a deployed system.