From Prediction to Decision in Financial Markets: Deep Learning in Market Heterogeneity and Multi-Agent Portfolio Optimization

Abstract

This thesis presents a systematic progression from predictive modelling to autonomous decision-making in financial markets, unified by the insight that financial time series are fundamentally distinguished from traditional domains by extreme asset heterogeneity, severe data sparsity, pronounced non-stationarity, and intricate inter-asset dependencies that cannot be adequately captured by historical price data alone.

The journey begins with the demonstration that aggregating stocks into universal numerical models neglects firm-specific characteristics and yields poor generalisation. A hybrid BiLSTM-Transformer architecture is proposed that incorporates BERT-encoded business descriptions capturing unique attributes such as industry sector and operational scale combined with a series of technical indicators and dimensionality reduction via Restricted Boltzmann Machines. This framework achieves superior accuracy and robust performance on previously unseen companies, establishing that effective stock prediction requires explicit modelling of corporate heterogeneity.

Building upon this foundation, the distinctive constraints of financial seq2seq forecasting are addressed through a novel patch-based Mamba-CrossAttention Network. By combining residual-connected Mamba blocks with localized patch processing and adaptive cross-layer attention, the architecture efficiently extracts temporal dynamics under conditions of extreme volatility and low signal-to-noise ratios, delivering state-of-the-art results across major global stock indices.

The analysis then extends to portfolio construction, revealing that isolated stock forecasts ignore systemic relational structures. A dual-graph BiLSTM-GAT framework is introduced, simultaneously learning technical similarity from price co-movements and fundamental affinity from industry relationships, with alignment performed via an attention mechanism.