llm-finance-agentslisted

# LLM trading agents — the evidence ## 1. The executive answer **No credible evidence exists that an LLM trading agent produces risk-adjusted alpha net of costs.** The literature splits cleanly: | Claim | Status | |---|---| | LLMs extract *sentiment* from news better than lexicon methods | ✅ well supported | | LLM sentiment scores correlate with subsequent returns in-sample | ✅ supported, **but heavily contaminated by lookahead** | | That correlation survives realistic transaction costs | ❌ **fails at ~20 bps round-trip** — by Lopez-Lira & Tang's own numbers | | Multi-agent debate beats one well-prompted agent | ❌ **not established**; the general MAD literature finds the opposite | | Reported Sharpe ratios of 5–8 are real | ❌ artifacts of 3-month windows, 3 tickers, zero costs, a bull regime | | LLM agents beat buy-and-hold in a contamination-free setting | ❌ margin inside noise; **in the one real-money competition, 4 of 6 models lost money** | **The one well-designed study kills its own strategy.** Lopez-Lira & Tang (arXiv 2304.07619) is genuinely post-cutoff and reports a gross Sharpe of 2.97. Its cost curve: **~700% at 0 bps → >300% at 5 bps → >100% at 10 bps → unprofitable at 20 bps round-trip.** That curve, not a point estimate, is the format this field should have adopted. **Alpha Arena S1** (Oct 18 – Nov 3 2025, **real capital**): 4 of 6 models lost money; GPT-5 finished **−59%**; fees alone ate **$1,654 of Qwen3's $10k**. 🚨 The widely circulated **"+79%" is a mi