Inside the Trading Loop: Markets, Models, and Moves with LLM Agents (arXiv (Cornell University))

Summary

• The higher paper account return was accompanied by a larger maximum drawdown of 26.8%, compared with 6.3% for SPY.p.7
• Reconstructed same-session portfolio replays across all configurations generated returns between 55.5% and 58.7%, compared with 14.8% for SPY.p.8
• Next-session assessment replays produced returns ranging from 71.0% (L1) to 78.3% (L2), outperforming matched SPY (14.8%) and QQQ (19.1%) references.p.37
• Enabling LLM sentiment inputs alone produced minimal allocation impact, with a mean daily allocation distance of 0.10 percentage points from the L0 baseline.p.24
• A margin-based call gate using dt < 4 selected 66% of run dates while capturing 98% of the advisory arm's allocation distance versus L0.p.28

Key insights

1. The raw return advantage of the paper trading account comes at the expense of substantially higher risk and drawdown compared to broad market benchmarks.p.7
2. Delegating macro stance directly to cash authority shifts cash exposure to 5% on almost every date, behaving similarly to a constant 5% cash control.p.9
3. Selectively calling the LLM based on macro-score distance to cash boundaries captures most allocation changes while reducing call volume.p.10
4. Shortening the response contract improves JSON parsing reliability on rich headline inputs but alters the macro stance recommendations.p.11
5. Model placement after top-k stock selection limits sentiment analysis to weighting adjustments without affecting portfolio membership.p.26
6. Lagging analyst assessments by one trading session alters holdings paths and assessment age, producing higher reconstructed returns across all tested configurations.p.8