Метод прогнозування криз фінансових ринків на основі векторної бази даних історичних подій та семантичного аналізу з використанням LLM

Abstract

This work is devoted to the development of a method for forecasting financial market crises based on a vector database of historical events. The proposed method employs a Multi-Factor Financial Foundation Model (MFFM) that fuses quantitative market indicators with semantic news embeddings into a unified vector representation. Unlike prior ML approaches to crisis prediction that rely on fixed feature sets and lack interpretability, the proposed method combines multimodal embeddings with retrieval-based reasoning. Walk-forward backtesting on data spanning 2000–2025 is used to evaluate the method against established baselines including VIX threshold, logistic regression, turbulence index, and absorption ratio.