Paper: anonymized LLM optimizers
A new arXiv discussion called 'Blindfolded LLMs' explores anonymized portfolio optimization that blends LLM techniques with traditional quant methods — a signal that large models are being tested for portfolio construction while researchers wrestle with data privacy and explainability. This is emerging research quants should watch for novel forecasting approaches and validation challenges. (x.com)
Paper authored by Joohyoung Jeon and Hongchul Lee, submitted to arXiv as arXiv:2603.17692 on 18 March 2026 and listed as accepted to the ICLR 2026 Workshop on Advances in Financial AI (FinAI). (arxiv.org) The authors present a framework called BlindTrade that anonymizes S&P 500 constituents daily, runs four specialized LLM agents (Momentum, News-Event, Mean-Reversion, Risk-Regime), aggregates reasoning embeddings with a Semantic Graph Encoder (SemGAT), and executes trades via a PPO-DSR reinforcement-learning policy. (arxiv.org) Reported out-of-sample performance for 2025 YTD (through 2025-08-01) shows an annualized Sharpe of 1.40 ± 0.22 across 20 random seeds, with the paper noting the policy beat SPY in all 20 seeds during that window. (arxiv.org) The experimental design explicitly targets two sources of spurious backtest gains—memorization from ticker-specific pretraining and survivorship/lookahead bias—using anonymization and negative-control tests to validate that signals are not driven by label leakage. (arxiv.org) Robustness checks extend the evaluation to the 2024–2025 period and report market-regime dependence: the learned policy outperforms in volatile regimes but shows reduced alpha in trending bull markets. (arxiv.org) One author lists Mirae Asset Securities as an affiliation alongside Korea University, signaling industry collaboration on the anonymization-first LLM-GNN-RL approach and its acceptance into an academic workshop focused on financial AI. (arxiv.org)
