When LLM Signals Hurt: A Coverage-Density Analysis of LLM-Augmented Reinforcement Learning for Stock Trading

We evaluate LLM-augmented reinforcement learning for stock trading on Nasdaq- 100 (2019–2023) and report a previously unmeasured experimental phenomenon: the relationship between LLM signal coverage density and trading performance is non-monotonic, with a clearly identifiable harmful regime. In a controlled coverage sweep over {0%,5%, 20%, 50%, 80%, 100%}, signal injection at 5% and 20% coverage degrades performance below the no-signal baseline, becoming net-positive only at ≥ 50% coverage. The FNSPID dataset’s 9.7% non-neutral coverage sits inside this harmful regime—meaning that for typical research configurations available today, adding LLM signals to the RL pipeline reduces returns. Beyond this density finding, we report three further negative results that the LLMRL trading literature has not adequately addressed. First, our LLM-augmented RL agent (158.11% cumulative return as a 3-seed ensemble) is outperformed by three standard non-RL baselines that prior work in this thread does not report: momentum top-10 (250.45%), equal-weight buy-and-hold (235.00%), and equal-weight monthly rebalanced (214.06%), all of which also exceed the Nasdaq- 100 buy-and-hold benchmark (164.52%). Second, we control for the daily-vs.- monthly rebalancing-frequency confound by deploying the same trained agents under matched-frequency monthly execution; the monthly variant underperforms its daily counterpart by 47pp (111.01% vs. 158.11%), confirming that the baseline gap is not driven by transaction-cost differences. Third, a v3-matched ablation finds that removing the CVaR tail-risk constraint produces a difference within the seedto- seed variability of the experiment. Across two independent runs, the sign of this difference flipped, providing direct empirical evidence that the algorithmic risk-tail machinery contributes no detectable return benefit in this setting. A regime decomposition reveals one clear win for the agent: in the 2023 recovery period, the 3-seed ensemble (52.6%) outperforms all non-RL baselines, suggesting the learned policy may have regime-specific advantages that single-window evaluation obscures. We argue that LLM-RL trading research should adopt non-RL baselines as standard practice, report signal coverage density as a first-class experimental variable, and decompose results by regime. Code and trained models are available at https: //anonymous.4open.science/r/signal-density-llm-trading-9966/.