RetailAgent: Structured Adverse Timing in Self-Conditioned Multimodal LLM Trading Agents

Battle Count: 72

Highly relevant to quantitative trading as it directly studies the behavioral structure of LLM-based trading agents, revealing persistent adverse timing (systematically wrong-signed entry/exit) that could be exploited by other market participants. The exposure-matched timing framework, trajectory shuffling controls, and signal-overlap tests provide methodological tools for diagnosing predictable policy structure. The self-conditioning amplification findings are relevant for designing LLM trading agents with memory. However, the controlled setting without transaction costs, market impact, or interactive feedback limits direct applicability to live trading systems.

Moderate-to-high complexity. Requires: multimodal LLM inference (text + vision pathways via vLLM), sequential decision loops over intraday grids, VICReg-pretrained GRU encoder with low-rank adapter for soft-token conditioning, two-pass narrative memory protocol, multiple stateful prompt variants (position, ledger, account), trajectory shuffling and complementary schedule diagnostics, signal-overlap projections, date-clustered bootstrap inference, and careful parsing/generation controls. The framework spans 14+ standard configurations plus narrative and latent-state ablations. However, the core LLM is frozen (no fine-tuning), and the encoder training is relatively lightweight (54k-parameter adapter).

The paper provides extensive reproducibility documentation including: recorded seeds for panel construction and evaluation, complete prompt templates, generation and parsing rules, saved trajectories with actions and memory excerpts, encoder checkpoints, failure counters, machine-readable evaluation outputs, and detailed computational resource records. The accompanying artifact contains panel-construction and scoring code. However, exact software versions, immutable model revision hashes, and full-project compute accounting are noted as items for final release. The coarse-image checkpoint identity has an author-verification ambiguity.