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

Joohyoung Jeon; Hongchul Lee

cs.LGMar 18, 2026

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

Authors: Joohyoung Jeon, Hongchul Lee

Organizations: Korea University · Mirae Asset Securities

Abstract

For LLM trading agents to be genuinely trustworthy, they must demonstrate understanding of market dynamics rather than exploitation of memorized ticker associations. Building responsible multi-agent systems demands rigorous signal validation: proving that predictions reflect legitimate patterns, not pre-trained recall. We address two sources of spurious performance: memorization bias from ticker-specific pre-training, and survivorship bias from flawed backtesting. Our approach is to blindfold the agents--anonymizing all identifiers--and verify whether meaningful signals persist. BlindTrade anonymizes tickers and company names, and four LLM agents output scores along with reasoning. We construct a GNN graph from reasoning embeddings and trade using PPO-DSR policy. On 2025 YTD (through 2025-08-01), we achieved Sharpe 1.40 +/- 0.22 across 20 seeds and validated signal legitimacy through negative control experiments. To assess robustness beyond a single OOS window, we additionally evaluate an extended period (2024--2025), revealing market-regime dependency: the policy excels in volatile conditions but shows reduced alpha in trending bull markets.

Figures & tables

Agent	RAW IC ( $p$ )	LLM IC ( $p$ )	$\Delta$ IC
Momentum agent	$-0.019$ ( $5{\times}10^{-5}$ )	$+0.001$ (0.58)	$+0.020$
News-Event agent	$+0.003$ (0.12)	$+0.006$ ( $3{\times}10^{-4}$ )	$+0.003$
Mean-Reversion agent	$-0.005$ (0.26)	$-0.000$ (0.97)	$+0.005$
Risk-Regime agent	$-0.006$ (0.27)	$+0.011$ ( $1{\times}10^{-4}$ )	$+0.017$
Average	$-0.007$	$+0.005$	$+0.011$

Table 1: LLM vs RAW Feature IC Comparison (h=21 days). Negative RAW IC indicates inverse correlation;

\Delta

IC toward zero can remove misleading signals rather than adding predictive power. We therefore report both absolute IC and

\Delta

IC.

Figure 1: Cumulative returns for 2025YTD OOS. Shaded band shows

\pm

1 std across 20 seeds.

Strategy	Sharpe (ann.)	CumRet (%)	MDD (%)	Vol (ann.)
BlindTrade	$\mathbf{1.40\pm 0.22}$	$\mathbf{32.22\pm 5.21}$	$-31.66\pm 5.76$	$42.34\pm 7.12$ %
SPY	0.64	8.52	$-19.00$	23.26%
EQWL	0.74	7.23	$-15.39$	22.51%
Momentum	0.89	15.42	$-26.49$	32.19%
MCap Top-20	0.85	11.50	$-19.56$	24.91%
RAW Top-20	0.97	11.54	$-15.63$	21.78%

Table 2: 2025YTD OOS Performance (2025-01-02 to 2025-08-01, 145 trading days). Baseline parameters: reward_cost_scale (

c

) = 0.358, rebalance_eta (

\eta

) = 0.10. Mean

\pm

std across 20 seeds. Sharpe and Vol are annualized. Cost: 10 bps/turnover. Abbreviations: SPY = S&P 500 ETF, EQWL = Equal-Weight ETF, CumRet = Cumulative Return, MDD = Maximum Drawdown, Vol = Volatility.

Variant	$\Delta$ Sharpe vs SPY	Std	Win Rate
SemGAT (baseline)	$+0.76$	$0.20$	20/20
SemGAT-C (Vol-target)	$+0.37$	$0.69$	12/20
SemGAT-D (Conf-loss)	$+0.16$	$0.40$	10/20

Table 3: 20-seed stability analysis. Sharpe difference vs SPY and win rate.

Appendix figures & tables12 assets

Supplementary material from the paper’s appendix.

Appendix

RAW Features (Input)	LLM Features (Output)	Used
price_vs_ma20: Price relative to 20-day MA	score: Momentum score ( $-1$ to $+1$ )	✓
price_vs_ma60: Price relative to 60-day MA	confidence: Agent confidence (0 to 1)	✓
adx_14: 14-day ADX (trend strength)	win_probability: Estimated win prob	✓
macd_histogram: MACD histogram value	momentum_stage: nascent/mid/late/exhausted	✓
rsi_14: 14-day RSI	reasoning: Free-text $\rightarrow$ embedding	✓
volume_vs_ma20: Volume relative to 20-day avg	trend_quality, sustainability_score	–

Appendix

Table 4: Momentum Agent: RAW (input) vs LLM (output) features.

RAW Features (Input)	LLM Features (Output)	Used
headlines: Anonymized news headlines	event_strength: Impact magnitude (0 to 1)	✓
sentiment_polarity: Pre-computed polarity	urgency: Time sensitivity (0 to 1)	✓
news_count_today: Number of headlines	primary_event_type: Event category	–
source_diversity: Unique sources count	attention_score, relative_attention	–

Appendix

Table 5: News-Event Agent: RAW (input) vs LLM (output) features.

RAW Features (Input)	LLM Features (Output)	Used
rsi_14: 14-day RSI	reversion_score: Signal ( $-1$ to $+1$ )	✓
bb_position: Bollinger Band position	confidence: Agent confidence (0 to 1)	✓
deviation_from_20ma: Price deviation	market_psychology: panic/euphoric/normal	✓
stochastic_k: Stochastic oscillator	extreme_level, reversal_probability	–

Appendix

Table 6: Mean-Reversion Agent: RAW (input) vs LLM (output) features.

RAW Features (Input)	LLM Features (Output)	Used
beta_sp500: Market beta	systemic_risk_score: Risk level (0 to 1)	✓
beta_vix: VIX sensitivity	confidence: Agent confidence (0 to 1)	✓
tail_beta: Tail risk beta	regime_class: defensive/neutral/cyclical/crisis	✓
var_95/cvar_95: Value-at-Risk metrics	reasoning: Free-text $\rightarrow$ embedding	✓
macro_context: Current regime input	crisis_sensitivity, regime_dependency	–

Appendix

Table 7: Risk-Regime Agent: RAW (input) vs LLM (output) features.

Metric	SPY	EQWL	Momentum	MCap Top-20	RAW Top-20	BlindTrade
	Passive BM		Active BM			Ours
Sharpe	1.06	0.66	1.34	1.54	1.69	$0.69\pm 0.23$
CumRet (%)	31.5	16.5	73.5	53.4	63.3	$30.1\pm 5.4$
AnnRet (%)	18.3	10.2	41.9	31.2	36.5	$21.8\pm 16.9$
Vol (%)	17.2	17.0	29.5	18.7	19.6	$36.2\pm 2.4$
MDD (%)	$-$ 19.0	$-$ 20.7	$-$ 26.5	$-$ 19.6	$-$ 15.6	$-32.1\pm 2.5$

Appendix

Table 8: Extended OOS Performance (2024-01-02 to 2025-08-01, 397 trading days). Parameters:

\eta

=0.08, cost_scale=0.40. Mean

\pm

std across 20 seeds. All metrics annualized.

Period	Strategy	Sharpe	CumRet (%)	AnnRet (%)	Vol (%)	MDD (%)
2024 (Bull)	SPY	1.70	23.3	23.3	12.4	$-$ 8.4
2024 (Bull)	BlindTrade	$0.34\pm 0.18$	$5.8\pm 5.8$	5.8	$29.6\pm 1.1$	$-18.5\pm 3.1$
2025 YTD (Volatile)	SPY	0.54	6.1	10.8	23.2	$-$ 19.0
2025 YTD (Volatile)	BlindTrade	$\mathbf{1.02\pm 0.12}$	$\mathbf{22.8\pm 4.0}$	46.7	$44.2\pm 1.2$	$-32.4\pm 0.7$

Appendix

Table 9: Performance breakdown by market regime. Mean

\pm

std across 20 seeds for BlindTrade. 2024 was a strong bull market; 2025 YTD exhibited high volatility.

Figure 2: The BlindTrade Pipeline: Data anonymization, Multi-agent LLM feature generation, IC validation, SemGAT encoding, Intent-conditioned RL (PPO-DSR), and Backtesting.

Figure 3: Leakage audit via negative control. When predictions are randomized, IC disappears and performance collapses.

Figure 4: Intent probability timeline across Train/Val/OOS periods. (a) Daily intent probabilities show how the policy adapts to market conditions. (b) Intent distribution remains stable across splits, demonstrating generalization.

Figure 5: Intent-conditioned policy behavior. (a) Defensive mode shows higher turnover (2.9%/day) for active rebalancing. (b-c) Max weight and concentration (Effective N) differ significantly by intent (Kruskal p=0.000).

Figure 6: t-SNE visualization of RL state vectors across Train/Val/OOS splits. Overlapping distributions suggest similar feature characteristics across splits.

Figure 7: RL Policy Network Architecture. The Intent Head selects defensive/neutral/aggressive mode from aggregated LLM statistics and GNN market state, then the Node Score Head produces stock-level scores, which parameterize a Dirichlet distribution for portfolio weights.

Explore similar work

From Knowing to Doing: A Memory-Controlled Benchmark for LLM Trading Agents on Stock Markets
May 27, 2026Taojie Zhu, Wentao Zhao, Rui Sun +7Benchmark Contamination LLM Agent Evaluation
1Tsinghua University, 2Stepfun · 2Stepfun · 4Shanghai Jiao Tong University +3
TrustTrade: Human-Inspired Selective Consensus Reduces Decision Uncertainty in LLM Trading Agents
Mar 23, 2026Minghan Li, Rachel Gonsalves, Weiyue Li +2Algorithmic Trading LLM Agent Reliability
Harvard AI and Robotics Lab, Harvard University . · Harvard Business School, Harvard University . · Daegu Gyeongbuk Institute of Science and Technology . +1
Agentic Trading: When LLM Agents Meet Financial Markets
May 19, 2026Yihan Xia, Panpan You, Taotao Wang +4LLM Agent Evaluation Finance
College of Electronic and Information Engineering, Shenzhen University, Shenzhen, China

May 27, 2026J/K move · Enter open · S save

From Knowing to Doing: A Memory-Controlled Benchmark for LLM Trading Agents on Stock Markets

Taojie Zhu, Wentao Zhao, Rui Sun, Beidi Luan+6

1Tsinghua University, 2Stepfun · 2Stepfun · 4Shanghai Jiao Tong University+3

Benchmark Contamination LLM Agent Evaluation AI Agent Benchmarks Finance Algorithmic Trading

Open paper AskPDF arXiv

Evaluating whether large language model (LLM) agents can profit in capital markets is increasingly framed as end-to-end trading: place an agent in a historical market, let it trade, and measure portfolio returns. This setup is vulnerable to two evaluation failures. First, long backtests often overlap with the knowledge cutoffs of frontier LLMs, allowing memorized tickers, dates, prices, and market narratives to substitute for investment reasoning. Second, raw returns are a noisy proxy for stock-selection ability, since positive performance may come from market beta, style exposure, or favorable regimes rather than genuine alpha. We introduce KTD-Fin (Knowing-To-Doing Financial Benchmark), an end-to-end stock-market trading benchmark that addresses both issues. KTD-Fin uses a data-side masking protocol to anonymize key identifiers and calendar information consistently across prompts and tools, separating historical market memory from investment decision-making. It also incorporates a Barra-style performance attribution framework that decomposes portfolio returns into market, style, and stock-selection alpha components. Across ten frontier LLM agents evaluated on the Chinese CSI300 over a 2024--2026 window, masking substantially changes agent rationales, pushing them towards anonymized factor-based reasoning. Attribution analysis further shows that LLM agents' cumulative returns under leakage-controlled evaluation are largely explained by passive market and style exposure, with limited evidence of persistent stock-selection alpha. These findings suggest that financial LLM benchmarks should evaluate not only whether an agent makes money, but also whether the source of returns reflects transferable investment skill. We release KTD-Fin as a reproducible template for leakage-controlled and attribution-aware evaluation of LLM trading agents.

Abstract

Explore similar work

From Knowing to Doing: A Memory-Controlled Benchmark for LLM Trading Agents on Stock Markets

TrustTrade: Human-Inspired Selective Consensus Reduces Decision Uncertainty in LLM Trading Agents

Agentic Trading: When LLM Agents Meet Financial Markets