Bayesian Change Point Estimation and Evaluation of Machine Learning Models for Multi‐Scale <scp>SPI</scp> Prediction for Assam‐Meghalaya Region in North‐Eastern India

ABSTRACT The study aimed to (i) identify breakpoints in rainfall patterns and (ii) evaluate the out‐of‐sample prediction accuracy of various machine learning models. Monthly rainfall data from 1901 to 2017 were obtained from the Government of India database. The methodology was based on SPI, Bayesian breakpoint detection and application of various machine learning models. The Diebold–Mariano test was also used to compare the performance of the models via accuracy metrics (root mean square error—RMSE, mean absolute error—MAE and mean absolute percentage error—MAPE). The division of SPI into three time periods (1901–1950, 1951–2000, 2001–2017) revealed more frequent rainy spells in the quarterly SPI (SPI3) compared to the other intervals (very wet conditions; 1st Interval: 30, 2nd Interval: 29, 3rd interval 6). On the other hand, the half‐yearly SPI (SPI6) showed an increase in dry spells during the second and third intervals. The nine‐month SPI (SPI9) exhibited similar patterns of drought periods to those of SPI3 and SPI 6, with severe droughts starting in 1975 in SPI9 and the annual SPI. Bayesian analysis identified breakpoints in SPI3 (18), SPI6 (23), SPI9 (26) and SPI12 (22) under threshold value 0.5 with maximum posterior probability and found an increase in the frequency of drought periods starting in 1975 mainly in the second and third intervals for SPI6, SPI9 and SPI12. The negative mean values of the SPI during these later intervals highlighted the intensity of the droughts. Regarding the model, specific Neural Network Autoregressive Models (NNAR) demonstrated the lowest estimated variance for each SPI time scale (SPI3‐0.263, SPI6‐0.230, SPI9‐0.198 and SPI12‐1.191). NNAR models consistently outperformed other machine learning techniques in terms of prediction accuracy and provided a better explanation of the SPI variability in all considered intervals. The Diebold–Mariano test confirmed better predictive accuracy for NNAR models for the SPI in the Assam‐Meghalaya region in northeastern India.