Learning to Aggregate Zero-Shot LLM Agents for Corporate Disclosure Classification

Battle Count: 174

Directly addresses predicting next-day stock return direction from corporate disclosures, a core input for event-driven trading strategies. The multi-agent zero-shot framework offers a practical, low-annotation-cost pipeline for financial text classification. The finding that aggregation helps most in mixed-signal disclosures (beat + weak guidance, growth + legal risk) is highly relevant for trading around earnings and regulatory filings. However, the binary next-day direction target is simplistic compared to full trading strategy backtesting, and no transaction costs, slippage, or portfolio construction are discussed.

Moderate complexity. Requires running three zero-shot LLM agents with fixed prompts (using open-source models like Qwen2.5 and Llama-3.2), extracting structured JSON outputs, computing token-level confidence, building a feature vector, and training a standard L2-regularized logistic regression. No fine-tuning of base models is needed. The main engineering effort is in the inference pipeline, caching, and preprocessing. The logistic meta-classifier itself is trivial to implement.

The paper describes a fully deterministic pipeline: fixed prompts, temperature=0, top-p=1, fixed random seed, stored model outputs, cached JSON results, L2-regularized logistic regression with liblinear solver, chronological train/dev/test split. All stages are deterministic conditional on stored model version, prompt text, and seed. However, no explicit GitHub repository or code link is provided in the extract.