Can Blindfolded LLMs Still Trade? An Anonymization-First Framework for Portfolio Optimization

1. Problem Statement

The paper addresses two critical flaws in current Large Language Model (LLM) trading agents:

• Memorization Bias: LLMs often rely on pre-trained associations with specific tickers (e.g., "Tesla" = "growth") rather than understanding underlying market dynamics. This leads to spurious performance that fails in real-world deployment.
• Backtesting Biases: Traditional financial backtesting suffers from survivorship bias (excluding delisted companies) and lookahead bias (using future data). These issues create unrealistic performance metrics that do not translate to live trading.

The core research question is: Can LLMs generate genuine predictive signals for portfolio optimization if all specific identifiers (tickers, company names) are removed to force reliance on market dynamics rather than memorization?

2. Methodology: The BlindTrade Framework

The authors propose BlindTrade, an end-to-end framework integrating Anonymization, Multi-Agent LLMs, Graph Neural Networks (GNN), and Reinforcement Learning (RL).

A. Data Anonymization (The "Blindfold")

• Ticker Masking: All stock tickers and company names are replaced with synthetic identifiers (e.g., "AAPL" $\to$ "STOCK_0026").
• Entity Masking: Proper nouns in news headlines (e.g., "Tim Cook," "iPhone") are anonymized using the Google Knowledge Graph API.
• Point-in-Time Universe: The system uses only the actual S&P 500 constituents for each specific date to eliminate survivorship bias.

B. Multi-Agent LLM Feature Generation

Four specialized LLM agents evaluate stocks based on anonymized data and technical indicators, operating under a strict knowledge cutoff (only data up to $t-1$ is visible):

1. Momentum Agent: Analyzes price trends and volume.
2. News-Event Agent: Evaluates sentiment and event impact from anonymized headlines.
3. Mean-Reversion Agent: Identifies overbought/oversold conditions.
4. Risk-Regime Agent: Assesses systemic risk and market regime.

• Output: Each agent outputs a score, confidence level, and free-text reasoning. The reasoning text is embedded into a 384-dimensional vector using SentenceTransformer.

C. Semantic Graph Encoder (SemGAT)

Instead of treating stocks independently, the framework constructs a dynamic graph to learn inter-stock relationships:

• Nodes: Represented by a 394-dimensional vector (LLM scores + reasoning embeddings).
• Edges:
  • Sector Edges: Connect stocks within the same GICS sector.
  • Semantic Edges: Connect stocks if the cosine similarity of their reasoning embeddings exceeds a threshold (0.75), allowing the model to learn "similar reasoning patterns" even without knowing the company names.
• Architecture: A 2-layer GATv2 encoder predicts return distributions using a HL-Gauss (Horseshoe-Gaussian) distribution loss combined with pairwise ranking loss.

D. RL Policy (PPO-DSR)

A Proximal Policy Optimization (PPO) agent determines portfolio weights based on the GNN outputs:

• Intent Head: Aggregates global statistics (momentum, risk, regime) to select a market posture: Defensive, Neutral, or Aggressive.
• Action Space: Allocates weights to the top-20 stocks (Top-K masking) within the S&P 500 universe. No cash allocation is permitted (100% equity exposure).
• Reward Function: Differential Sharpe Ratio (DSR), penalizing volatility and transaction costs (10bps).
• Execution Inertia: A parameter ($\eta$) controls turnover to prevent excessive rebalancing.

3. Key Contributions

1. Anonymization-First Protocol: A rigorous method to block memorization by replacing all identifiers, forcing LLMs to rely on structural market patterns.
2. Semantic Graph Learning: A novel GNN approach that constructs edges based on the similarity of LLM reasoning rather than just static industry classifications, enabling relationship learning in an anonymized state.
3. Rigorous Signal Validation: The use of Information Coefficient (IC) analysis and Negative Control Experiments (randomly shuffling predictions) to prove that signals are not leakage artifacts.
4. Interpretable Intent: A mechanism that explicitly exposes the agent's market posture (Defensive/Neutral/Aggressive), bridging the gap between black-box RL and human interpretability.

4. Experimental Results

The model was evaluated on the S&P 500 constituents from 2020 to August 2025.

• Out-of-Sample (OOS) Performance (2025 YTD):
  • Sharpe Ratio: $1.40 \pm 0.22$ (significantly outperforming SPY at 0.64 and Momentum strategies at 0.89).
  • Cumulative Return: $32.22\% \pm 5.21\%$.
  • Volatility: $42.34\%$ (higher than SPY due to 100% equity exposure and active trading).
  • Max Drawdown: $-31.66\%$.
• Ablation Studies:
  • Removing LLM features dropped Sharpe from 1.40 to 1.14.
  • Removing the GNN structure caused a catastrophic drop to 0.62, proving that learning inter-stock relationships is critical.
  • Replacing RL with simple Top-K equal weighting resulted in a negative Sharpe (-1.17) due to high turnover costs, highlighting the necessity of cost-aware RL.
• Leakage Audit (Negative Control):
  • When prediction scores were randomly shuffled, the IC dropped to near zero ($0.0004$) and performance collapsed. This confirms the original signals contain legitimate predictive structure, not random noise or data leakage.
• Market Regime Dependency:
  • Volatile Markets (2025 YTD): BlindTrade excelled (Sharpe 1.02 vs SPY 0.54).
  • Bull Markets (2024): The strategy underperformed (Sharpe 0.34 vs SPY 1.70) because its diversified, defensive nature limited upside capture during sustained rallies.

5. Significance and Conclusion

BlindTrade demonstrates that LLMs can be trustworthy trading agents if they are forced to "think" without relying on memorized ticker associations.

• Trustworthiness: By proving that performance persists even when "blindfolded," the paper validates that the model learns genuine market dynamics.
• Validation First: The paper establishes a new standard for financial AI: signals must pass IC validation and negative control tests before deployment.
• Limitations: The strategy currently lacks a cash allocation mechanism, leading to higher drawdowns in crashes, and shows reduced efficacy in strong, sustained bull markets compared to passive benchmarks.

The work suggests a path forward for responsible AI in finance: Anonymization + Multi-Agent Reasoning + Graph Learning + Cost-Aware RL creates a robust, interpretable, and genuinely predictive trading system.