Interpretable Systematic Risk around the Clock

Battle Count: 184

Highly relevant to quantitative trading. The paper provides a real-time, implementable factor-mimicking portfolio strategy with out-of-sample Sharpe ratio of 0.95 (vs. market 0.53), robust to transaction costs (SR 0.85-0.93 at 10-50 bps). The LLM-based jump classification enables interpretable risk decomposition and dynamic portfolio allocation. The around-the-clock analysis is particularly timely given NYSE/Nasdaq plans for 22-24 hour trading sessions. The methodology directly addresses risk management, portfolio construction, and real-time signal generation.

High complexity due to: (1) processing millisecond-level TAQ data for S&P 1500 constituents over 23 years; (2) constructing around-the-clock market factor combining intraday equity returns and overnight E-mini futures; (3) implementing continuous-time Fama-MacBeth regression with jump identification; (4) hosting and serving a 235B-parameter MoE LLM (requires 4x H100 80GB GPUs); (5) three-step prompting pipeline for narrative retrieval and classification; (6) real-time portfolio construction with annual rebalancing. The paper provides detailed code snippets for model serving and data processing.

The paper uses open-source models (Qwen3-235B-A22B, ChronoBERT) and provides detailed hosting instructions with vLLM. Data sources are identified (TAQ, CRSP, Dow Jones Newswires, CME DataMine, Compustat). However, Dow Jones Newswires data is proprietary and may limit full replication. The paper provides extensive appendix details on data processing, model serving, and fine-tuning procedures. No GitHub repository is explicitly mentioned.