Exploring SPI-Based Machine Learning and Statistical Models for Drought Prediction in Southeastern Romania

Abstract

This study analyzes the performance of several forecasting models applied to Standardized Precipitation Index (SPI) time series for the 1967–2021 period, using data from six weather stations in southeastern Romania. Four temporal aggregation scales (SPI3, SPI6, SPI12, and SPI24) were considered for one-step-ahead (one-month) forecasting. The evaluated models included Random Forest, XGBoost, Support Vector Regression (SVR), N-BEATS (generic variant), and SARIMA (Seasonal Autoregressive Integrated Moving Average), covering machine learning, deep learning, and statistical approaches. To ensure a rigorous out-of-sample evaluation, SPI distribution parameters were estimated solely based on the pre-test calibration period and held constant when calculating SPI values for the independent test period, thereby preventing information leakage from the preprocessing stage. Forecast performance was assessed using RMSE, MAE, NSE, and a skill score relative to a persistence benchmark, while differences in forecast accuracy were statistically evaluated using the Diebold–Mariano test. The results revealed a clear scale-dependent pattern: absolute forecast accuracy increased with the SPI aggregation scale—alongside more pronounced temporal persistence—though this did not necessarily translate into superior skill compared to the persistence model. Among the data-driven models, SVR demonstrated the most robust performance, particularly at shorter timescales, while N-BEATS also provided competitive forecasts without consistently outperforming SVR. SARIMA showed the clearest gains over the persistence model at the SPI12 and SPI24 scales. The maximum NSE increased from 0.61 for SPI3 and 0.81 for SPI6, both obtained with SVR, to 0.94 for SPI12 (obtained with SARIMA) and 0.97 for SPI24 (obtained with SARIMA, N-BEATS and SVR). In conclusion, the results demonstrate that the SPI accumulation scale, temporal structure, and performance relative to persistence must be considered simultaneously when selecting forecasting models for drought prediction.