Financial prediction models performed close to chance when researchers removed an information-leakage problem. A controlled simulation also showed that the models could detect a deliberately introduced signal.
What researchers studied
The team designed an audit of machine-learning methods used to forecast financial returns, focusing on the risk that models accidentally access information unavailable at the prediction date.
The benchmark used 1,304 simulated business days, nine sector-specific asset series and seven economic indicators. A GARCH-like process generated volatility and deliberately controlled the relationships between inputs and future returns.
Researchers compared an LSTM neural network, a simplified attention model, XGBoost and a fixed-weight ensemble. The common sequential test subset contained 1,476 asset-date observations, ensuring comparisons used matching cases.
What they found
After correcting a setup where contemporaneous predictors were mechanically related to the target, average R-squared values ranged from -0.0052 to 0.0000 under the no-signal condition.
Directional accuracy ranged from 49.53% to 49.95%, effectively near the 50% benchmark. Bootstrap uncertainty intervals included chance performance for all evaluated models.
When the researchers deliberately introduced a future-return dependency involving changes in the VIX, the models recovered the known signal. At the strongest injected signal, average R-squared reached 0.9031 across the specified validation runs.
What it means
A sophisticated model can look impressive because of a faulty evaluation design rather than genuine predictive power.
Financial analysts should check exactly when each input becomes available, ensure that training does not use future information, and compare competing algorithms on identical test observations.
The study also demonstrates why explainability scores are insufficient on their own. A feature may appear influential without being legitimately available at the moment a forecast would be made.
Important limitation
This was a deliberately controlled, synthetic financial benchmark rather than evidence of trading profitability in live markets.
The results do not prove that all financial AI models lack predictive power. They establish how these particular models behaved under known null and injected-signal conditions.
Real markets contain changing regimes, transaction costs and sources of data leakage that were not fully represented by this simulation.
Source Information
Study: Auditing information leakage and predictive signal in financial machine learning: a controlled study
Authors: Iván Ortiz-Gárces, Aracely Mera-Navarrete, Paul Sarango-Lalangui, Jefferson Beltrán-Morales and William Villegas-Ch
Journal / Source: Frontiers in Artificial Intelligence
Published: 12 November 2026
