Agentic Trading Systems: Do LLMs Justify Their Costs?

The increasing use of large language model (LLM) agents in trading systems raises a crucial question: do the costs associated with their reasoning, tool use, and continuous decision-making translate into measurable incremental profit? Existing evaluations typically focus on performance metrics, often overlooking the agentic viability—whether these dynamic, LLM-mediated decisions actually pay for their own intelligence. To address this, researchers developed TradeLens, a trace-grounded diagnostic toolkit. TradeLens analyzes trading records, runtime traces, and deployment configurations to reconstruct trading trajectories, attribute profit and cost to specific evidence, and diagnose why an agent either succeeds or fails to cover its intelligence costs. Extensive analysis across different backbone models, capital scales, trading frequencies, and system architectures revealed that viability fundamentally hinges on the intelligence-to-profit conversion rate. Various models exhibited distinct failure patterns, such as DeepSeek-V3.2 struggling with asset selection and GLM-4.7 showing negative timing. Factors like capital scale, trading frequency, and architecture primarily amplified or degraded the decision-attributed timing value, rather than being primary drivers of viability themselves. These findings shift the evaluation paradigm for LLM-based trading agents from mere capability ranking to a trace-grounded diagnosis of their economic intelligence.