When an LLM trades stocks profitably, is it reading the market or remembering the ticker? Language models trained on internet text have seen years of financial commentary linking specific tickers to outcomes. A model that “predicts” AAPL will rise might be recalling training data, not analyzing the current chart.
Jeon and Lee strip the identifiers. Ticker symbols and company names are anonymized before the data reaches the LLM agents. The models see price patterns, volume, and technical features but cannot connect them to any entity they might have memorized. Four LLM agents output scores and reasoning; a graph neural network built from their reasoning embeddings feeds a PPO-DSR policy optimizer.
The result: Sharpe ratio 1.40 on 2025 data, consistent across runs. The models can trade without knowing what they're trading. But extended testing reveals a limit: the approach outperforms in volatile markets and underperforms in trending bulls. The blindfolded trader excels at reading regime changes but misses momentum — the kind of signal that benefits from knowing which company is which.
The structural point: anonymization is not just a bias control; it's a capability test. By removing the possibility of memorization, the framework isolates what the model actually learned about market dynamics from what it memorized about specific assets. The Sharpe ratio under anonymization is the floor of genuine capability. Everything above it, in the non-anonymized case, is an unknown mixture of real skill and data leakage. The blindfold doesn't weaken the trader — it measures the trader.