HYBRID NEURAL NETWORK ARCHITECTURE FOR MODELING SOCIO-ECONOMIC SYSTEMS UNDER EXOGENOUS GEOPOLITICAL SHOCKS

Abstract

The article aims to provide theoretical justification and a detailed description of a conceptual hybrid neural network architecture that integrates Large Language Model (LLM) analysis, Long Short-Term Memory (LSTM) deep learning for time series, and Agent-Based Modeling (ABM) to improve the accuracy of identifying phase transitions in complex socio-economic systems (CSES) under conditions of military aggression and economic instability.

The research methodology is based on a synthesis of nonlinear dynamics, complex systems theory, and modern machine learning. The study applies theoretical analysis of scientific concepts — from B. Mandelbrot's stylized facts and S. Taylor's time series modeling to modern Transformer architectures — combined with mathematical formalization and architectural design of neural network systems. The agent-based modeling apparatus with an LLM-driven expectation formation function is employed to describe market participants' behavior.

A three-dimensional CSES modeling concept is proposed, comprising: (1) the perceptual level — cognitive extraction of geopolitical narratives via LLM with Chain-of-Thought prompting; (2) the analytical level — detection of recurrent time-series patterns through LSTM with forget and input gate mechanisms; (3) the behavioral level — simulation of emergent interaction among heterogeneous agents. The key contribution is the authors' synthesis operator Yₜ = f (TSₜ, Sentₜ, Behₜ; Θ), implemented as a Fusion layer — a fully connected neural network that dynamically weights the contributions of quantitative market indicators (TSₜ), cognitive sentiment (Sentₜ), and the behavioral vector (Behₜ). It is demonstrated that nonlinear aggregation of these components enables the model to replicate the resonance effect — when a geopolitical shock is amplified by agents' panic against a background of pre-existing high volatility — which is fundamentally impossible in traditional additive econometric models.

The paper introduces the first comprehensive architecture that, unlike existing approaches (ML ensembles, isolated LLMs, or standalone ABM), integrates all three paradigms at the level of vector embeddings rather than simple result averaging. This addresses a critical gap in the literature: the absence of an integrated framework for simultaneously incorporating 'hard' (price series), 'soft' (geopolitical narratives), and 'micro-behavioral' (agent emergence) data in identifying bifurcation points within CSES.

The proposed architecture enables the development of early-warning systems for financial regulators and institutional investors. The resulting CSES 'digital twin' allows for identification of critical system states prior to the onset of a crisis phase, thereby facilitating the design of preventive systemic risk management strategies under geopolitical turbulence, including the context of full-scale armed conflict and its associated global economic transformations.