Skip to main content
Back

BlindTrade Hides Stock Names from AI and Beats the S&P 500 by 2x

BlindTrade strips tickers and company names from LLM inputs to test whether trading agents understand markets or just memorize training data patterns.

Pro

Paper, Full Access
  • BlindTrade anonymizes tickers and company names before feeding data to four LLM trading agents.
  • Achieved annualized Sharpe of 1.40 versus SPY's 0.64 on 2025 YTD across 20 seeds.
  • Uses four specialized agents (momentum, news, mean-reversion, risk-regime) plus GNN and PPO policy.
  • Negative control shuffling dropped RankIC from 0.015 to 0.0004, confirming signal legitimacy.
  • Underperforms in bull markets (Sharpe 0.34 in 2024) but excels in volatile regimes.
  • Full prompts and hyperparameters released; datasets planned for release upon publication.

When you ask a large language model about Tesla or Nvidia, it doesn't process the request neutrally. It carries baggage from pre-training, including years of headlines like "Tesla stock surges" or "Apple crushes earnings." A new paper argues this is the wrong foundation for a trustworthy trading system, and proposes a simple fix: blindfold the model.

The framework, called BlindTrade, replaces every ticker and company name with anonymous IDs before letting four specialized LLM agents score stocks. Even without knowing which stock is which, the system hit an annualized Sharpe ratio of 1.40 during 2025 year to date, well above the S&P 500's 0.64 over the same window.

Ghosts in the training corpus

Recent work shows that LLMs exhibit pre-trained preferences for technology and large-cap stocks, leading to confirmation bias. That raises the question the authors want to answer: are LLMs actually understanding market dynamics, or pattern-matching against phrases they saw during pre-training?

The paper identifies two sources of fake performance that plague LLM trading research:

  • Memorization bias: The model recognizes "AAPL" and outputs whatever sentiment dominated its training corpus.
  • Survivorship bias: Backtests that use today's S&P 500 constituents ignore the fact that failed companies have already been removed, artificially inflating results.

Prior financial LLMs like FinBERT and BloombergGPT show tickers directly during training and evaluation, making it impossible to distinguish memorization from real understanding. Real-time evaluations have started exposing the gap. LiveTradeBench evaluated LLM agents in live environments and found that models which excelled on static benchmarks performed worse in actual trading.

Blindfolding the agents

BlindTrade's core trick is an anonymization protocol. Tickers get mapped to synthetic identifiers, so "AAPL" becomes "STOCK_0026", and proper nouns like "Apple", "iPhone", and "Tim Cook" in news are replaced using the Google Knowledge Graph API. The authors are careful not to overclaim. They do not argue this blocks all leakage, only that it blocks the path where an LLM sees a ticker and decides "it's Apple, so buy."