Large Language Model-Driven Small-Capitalization Trading: Integrating Financial News Sentiment, Macroeconomic Indicators, and Technical Signals

Battle Count: 56

Directly addresses quantitative trading strategy development for small-cap equities. Proposes a complete end-to-end pipeline from news sentiment extraction through portfolio construction. The decomposition of macro-beta triggers into pure alpha and pure beta legs is a novel contribution to stock selection methodology. The uncertainty-aware covariance feeding into allocators is a practical innovation. Results are evaluated under realistic transaction costs (up to 100 bps). The paper provides actionable insights on holding period selection, allocator choice, and regime decomposition for small-cap trading.

High complexity due to: (1) multimodal network with separate news and daily branches, (2) MC-Dropout for epistemic uncertainty with 50 stochastic passes, (3) rank-2 low-rank plus diagonal covariance parameterization, (4) agglomerative clustering for news deduplication, (5) entity-prior correction and trailing standardization for sentiment, (6) six different allocation methods, (7) three stock-selection regimes with rolling beta computations against 58 indicators, (8) log-to-arithmetic return conversion with positive-definite regularization, (9) transaction cost modeling with slippage. Requires expertise in NLP, deep learning, portfolio theory, and financial data engineering.

Code is available at an anonymous repository. The methodology is extensively detailed with formulas in appendices. However, the pipeline is complex with many hyperparameters (S=50 MC passes, dropout rate 0.2, clustering threshold 0.90, risk aversion 2.5, position cap 40%, etc.). The evaluation is limited to a single calendar year (2025) out-of-sample, and the large grid search across many configurations raises multiple-testing concerns. Fixed random seed of 7 is reported.