AN ADAPTIVE NEURAL-FUZZY MODEL FOR ASSESSING INVESTMENT RISKS IN THE AGRI-INDUSTRIAL COMPLEX: MODERNISATION OF THE DECISION-SUPPORT ARCHITECTURE THROUGH THE INTEGRATION OF MACHINE LEARNING

Abstract

The assessment of investment risks in the agro-industrial complex operates under conditions of structural uncertainty driven by climate volatility. Existing approaches — econometric models, machine-learning classifiers, and static fuzzy-logic frameworks — each sacrifice either interpretability or adaptability. This study proposes a three-level decision-support architecture in which the second-level fuzzification module is replaced by an adaptive neuro-fuzzy inference system (ANFIS). The ANFIS core trains on the Physical Output Index and related agro-industry indicators for Kazakhstan over 2010–2023, a period containing four confirmed drought years. Generalised Bell membership functions are calibrated via hybrid learning: least-squares estimation in the forward pass for consequent parameters, and gradient descent in the backward pass for antecedent parameters.

The results of a leave-one-out cross-validation experiment show that ANFIS-calibrated membership functions correctly classify all four drought years as "High Risk" and produce a temporal lead of up to two to three months relative to official statistics. The calibrated functions reveal an FDI saturation effect and a non-linear precipitation threshold, both of which the baseline symmetric triangular functions cannot capture. The Fuzzy-SAW aggregation with AHP weights is retained at Level 3, preserving the interpretability of the final risk score for investors and regulatory bodies. The proposed framework is positioned as the analytical core of an agricultural digital twin within the Agri-Food 4.0 paradigm.