Temporally Validated Rice Productivity Prediction in West Sumatra: Linear Regression Versus Random Forest

Rice productivity is an important indicator of food security in West Sumatra Province, shaped by climate and harvested area, so an accurate prediction model is needed. This study compares Linear Regression and Random Forest for predicting rice productivity using Area Sampling Frame (KSA) data, rainfall, and air temperature from 2020–2024, totaling 156 valid observations. Following CRISP-DM, inferential analysis used Multiple Linear Regression with assumption tests, while predictive evaluation used a temporal holdout split and pipeline preprocessing to prevent leakage. Features included lag-1 climate and productivity, region, and month; performance used R², RMSE, MAE, and MAPE against mean and seasonal-naive baselines. Inferential results show only air temperature significantly affects rice productivity (p=0.001), yet the model explains just 9.7% of variance (R²=0.097) and violates classical assumptions. An exploratory diagnostic identified an abrupt regime shift in 2024, which was excluded so models were trained and validated on the stable 2020–2023 period. Both models forecast future months well, with Linear Regression the best and most stable (holdout R²=0.70; rolling-origin R²=0.71±0.11) and Random Forest close behind (holdout R²=0.67). Lag-1 productivity dominated feature importance, indicating largely autoregressive accuracy, while climate and area add modest signal. Future work should add agronomic, remote-sensing variables and autocorrelation-aware models