BlindTrade proposes an anonymization-first framework for portfolio optimization that addresses two key sources of spurious performance in LLM-based trading: memorization bias from ticker-specific pre-training and survivorship bias from flawed backtesting. The framework anonymizes all stock identifiers (tickers, company names, product names) using synthetic IDs, then employs four specialized LLM agents (Momentum, News-Event, Mean-Reversion, Risk-Regime) to evaluate stocks from different perspectives. Each agent outputs scores and reasoning under strict knowledge cutoff and anonymized inputs. A Semantic Graph Encoder (SemGAT) constructs a dynamic graph using sector connections and reasoning embedding similarity, enabling inter-stock relationship learning. Finally, a PPO-DSR reinforcement learning policy determines portfolio weights with an intent mechanism that selects defensive/neutral/aggressive modes based on cross-sectional agent consensus.

On 2025 YTD S&P 500 data (through August 1, 2025), BlindTrade achieves an annualized Sharpe ratio of 1.40±0.22 and cumulative return of 32.22%±5.21% across 20 seeds, outperforming all benchmarks including SPY (Sharpe 0.64), equal-weight ETF (0.74), and momentum strategies (0.89). Ablation studies confirm that LLM features contribute +0.26 Sharpe, graph structure contributes +0.78, and RL cost-aware execution is essential for profitability. Negative control experiments (random shuffling of GNN scores) cause performance to collapse, validating that signals reflect genuine predictive structure rather than leakage. IC analysis shows that LLM agents improve predictive power over raw features, with the Risk-Regime agent showing the strongest improvement. The framework also demonstrates interpretability through intent modes and reasoning embeddings, though it exhibits market-regime dependency—excelling in volatile conditions but underperforming in strong bull markets. Overall, BlindTrade provides a rigorous methodology for building trustworthy LLM trading systems by ensuring predictions reflect legitimate market understanding rather than memorized associations.