SPEI drought forecasting in Sri Lanka

The question

How well can past observations predict next month’s drought index at three Sri Lankan locations? SPEI measures wet and dry conditions. This benchmark predicts SPEI-3 and SPEI-6, which describe the water balance over three- and six-month windows, one month ahead.

From the original study to a reproducible benchmark

My 2021 undergraduate research explored recurrent neural networks for drought prediction. The current repository rebuilds the analysis using recovered data from Buttala, Padaviya, and Tissamaharama. It contains data checks, forecasting code, model comparisons, saved predictions, and uncertainty estimates.

The evaluation moves forward through time: each model is selected using earlier data before being tested on the following period. Persistence, which repeats the latest observed value, provides a simple reference point.

Results

Across the 2000–2019 retrospective panel, ridge regression reduced mean absolute error for both targets. These scores average the three locations equally; lower values indicate more accurate predictions of the index.

What the comparison shows

Predicting the index more accurately did not always improve drought detection. At the drought threshold of SPEI ≤ −1, ridge improved the drought F1 score for SPEI-6 but scored below persistence for SPEI-3. F1 balances missed droughts and false alerts.

This is a historical benchmark on a previously inspected dataset. The recovered inputs do not fully reproduce the original four-region study, and their SPEI calibration details are incomplete. The results therefore describe the present three-location analysis; performance with newly arriving observations remains to be tested.