From Hypotheses to Factors: Constrained LLM Agents in Cryptocurrency Markets

Battle Count: 74

Highly relevant to quantitative trading. The paper directly addresses factor discovery for portfolio construction in cryptocurrency markets, provides concrete performance metrics (44.55% annualized return, Sharpe 1.55 OOS), includes transaction cost analysis, discusses capacity constraints, and presents a reproducible framework for systematic strategy development. The constrained DSL approach is directly applicable to production trading systems requiring auditability. The focus on out-of-sample validation and net-of-cost performance aligns with practitioner needs.

Moderate-to-high complexity. Requires: (1) LLM API integration with structured prompting, (2) implementation of a constrained factor DSL with expression tree parsing, (3) deterministic evaluation engine with cross-sectional IC computation, (4) portfolio construction and backtesting infrastructure, (5) append-only experiment logging system, (6) ridge regression for factor combination, (7) transaction cost modeling. The DSL design and evaluation engine are the most complex components. However, the framework avoids deep learning training, using only ridge regression for aggregation.

The paper emphasizes reproducibility through: (1) fixed data splits (2020-2022 train, 2023 validation, 2024+ OOS), (2) pre-specified selection gates with fixed thresholds, (3) append-only experiment traces recording all hypotheses, recipes, metrics, and decisions, (4) constrained DSL preventing arbitrary code generation, (5) deterministic evaluation engine. However, the specific GPT-5.4 prompts, exact gate thresholds, and full code are not explicitly provided in the extract. The DSL grammar is described but not fully formalized.