Hybrid Residual SARIMAX-XGBoost for Provincial Rice Production Forecasting in Indonesia: A Systematic Literature Review and Conceptual Framework for Food Policy Decision Support

Abstract

This study contributes a systematic literature review and conceptual framework for provincial rice production forecasting in Indonesia by integrating residual-based hybrid modelling, explainability, and food policy decision support. A total of 60 articles published between 2023 and 2026 were selected from 124 initial records identified from six academic search sources: ScienceDirect/Elsevier, IEEE Xplore, SpringerLink, MDPI journals, Scopus/Crossref tracing, and Google Scholar/manual tracking, using a PRISMA-style screening process. The review synthesizes evidence across rice production forecasting, crop-yield forecasting, hybrid machine learning, explainable artificial intelligence, and decision-support-oriented forecasting. The literature findings show that recent studies increasingly use agricultural statistics, climate variables, remote sensing, UAV imagery, crop-model outputs, machine learning, and deep learning. However, the reviewed studies rarely integrate province-aware determinants, residual-based model decomposition, time-aware validation, residual-level explainability, and DSS-ready policy outputs within a single framework. Based on this gap, the study proposes a conceptual Hybrid Residual SARIMAX-XGBoost framework. SARIMAX is positioned to model temporal, seasonal, autoregressive, and exogenous baseline patterns, while XGBoost learns nonlinear residual structures. SHAP is positioned as an explanation layer for interpreting forecast adjustments rather than as causal evidence. The proposed framework remains conceptual and requires future empirical validation using Indonesian provincial rice production data. This review provides theoretical contributions to hybrid forecasting literature and practical implications for information systems, decision support, and public-sector food policy planning.