Novel framework and reference architecture for artificial intelligence models for stock markets

Abstract

This original research article proposes, as its contribution, a novel unified framework and reference architecture for stock market prediction in postCOVID financial markets, which exhibit unprecedented volatility and nonlinear dynamics, demanding more robust predictive approaches than traditional models can provide. This original framework integrates artificial intelligence (AI) and machine learning (ML) models, ranging from classical techniques support vector machine (SVM) to deep learning (DL) architectures such as long short-term memory (LSTM) neural networks and gated recurrent unit (GRU) models, within a modular system encompassing data ingestion, sentiment processing, predictive optimization, reinforcement learning (RL), and cloud-based portfolio management. Another key original contribution is the synthesis of standards (ISO 23053, ISO 38505, ISO 20546) with big data methodological frameworks (REBD and Biggy), forming a unified meta-framework that orchestrates predictive signals from sentiment analysis (SA) and macroeconomic indicators. Experimental realworld stock market validation on mining-sector stocks demonstrated, with a 100% success rate, consistent investment outperformance over passive Buy & Hold baselines, yielding investment optimizations of up to +11.11 pp: the evaluated portfolios achieved 23.57% and 8.25% returns versus their 19.94% and 5.41% baselines, respectively. These results confirm the validity of the proposed novel framework as a reproducible reference architecture, an original contribution empirically grounded and experimentally validated for the development of future financial AI systems.