MemGuard-Alpha: Detecting and Filtering Memorization-Contaminated Signals in LLM-Based Financial Forecasting via Membership Inference and Cross-Model Disagreement
Roy and co-authors (2026) study news sentiment and text signals using a large language model. The abstract mentions a Sharpe ratio, alpha, an out-of-sample test and transaction costs but not a benchmark, multiple testing or the size of the universe.
“Evaluated across seven LLMs (124M to 7B parameters), 50 S&P 100 constituents, 42,800 prompts, and 5.5 years of daily data (January 2019 to June 2024), CMMD achieves a Sharpe ratio of 4.11 versus 2.76 for unfiltered LLM signals, a 49% improvement after transaction costs.”
StrategyNews sentiment
TechniqueLarge language model
FrequencyDaily
ClaimSharpe ratioalpha
Also stated: Text and news, out-of-sample test, costs
Missing from abstract: benchmark, multiple testing, universe
Institutional-grade audit of this strategy.
£5 — Request the audit- Audit of the paper claims accounting for costs, luck, crises, originality, geography, etc.
- Reproduction code integratable into Bloomberg, IBKR, etc.
- Bias-free datasets for backtesting included
- Delivered in 2 working days
The paper's neighbourhoodNews sentiment and text signals by year
FamilyNews sentiment and text signals148Momentum and trend following289Factor investing240Machine learning return prediction15318 more · 854
AssetEquities97
The family is counted across the whole universe. Asset, method and authors are counted inside the family.
| Similar papers | citation | same method | same asset | same authors | |
|---|---|---|---|---|---|
| MemGuard-Alpha: Limits of Membership Inference for Detecting and Filtering Memorization-Contaminated Signals in LLM-Based Financial Forecasting by Roy and Roy | 2026 | ||||
| Execution-Aware Alpha Mining: Teaching LLM Factor Agents to Account for Trading Costs by Nagireddy | 2026 | ||||
| Memorisation or Alpha? Detecting Look-Ahead Contamination in Cross-Sectional Equity Signals by Nguyen | 2026 | ||||
| Beta Recall, Alpha Recall, and a Contamination Detector that Needs No Labels * Measuring Training-data Leakage in LLM Equity Signals by Nguyen | 2026 | ||||
| Crude oil arbitrage and momentum trading strategies with targeting volatility by using Large Language Models by Garmash | 2026 |