Forecasting Rice Production in Northern Samar: A Multi-Model Analysis Using Cyclone Frequency and Harvested Area as Predictors

Abstract: This study forecasted rice production in Northern Samar using three predictive models: Multiple Linear Regression (MLR), Random Forest Regression (RF), and Seasonal Autoregressive Integrated Moving Average with Exogenous Variables (SARIMAX). The analysis utilized cleaned datasets from 2016 to 2024, covering 23 municipalities after excluding San Vicente due to data insufficiency. Two explanatory variables—area harvested and tropical cyclone frequency—were used, with the former projected through Random Forest (SMAPE = 12.75%) and the latter using a constant mean approach. For rice production forecasts, MLR demonstrated steady upward trends and yielded the lowest MAPE in 4 municipalities, notably Catarman (5.48%) and Las Navas (3.91%). RF achieved the lowest MAPE in 17 municipalities, with exceptional accuracy in Lapinig (1.63%), Laoang (3.58%), and Lavezares (3.40%); however, it failed to capture long-term growth as it produced identical forecasts for 2030, 2035, and 2040—indicating limitations in trend sensitivity due to capped inputs. SARIMAX, while incorporating time-series structure and external factors, displayed inconsistent performance with the lowest MAPE in only 2 municipalities (Allen: 6.56%, Biri: 25.43%). Its trend outputs were often unstable in areas with irregular patterns, making it less suited for long-range forecasting. Among the three models, MLR proved most practical for long-term planning due to its interpretability and ability to reflect realistic growth. RF was best suited for short-term accuracy and baseline projections, while SARIMAX offered mixed results depending on data structure. These forecasts serve as valuable tools for local agricultural planning, guiding food security initiatives and resource allocation in Northern Samar.