Do genomic AI models actually learn anything useful?

Published at ICLR 2026, this paper from M42 and ADIA Lab tests seven Genomic Foundation Models (GFMs) across 52 tasks — comparing them against models with completely random, untrained weights. The results are uncomfortable: randomly initialised models are surprisingly competitive, and the biggest performance driver turns out to be tokenisation strategy, not pretraining. More critically, all evaluated models show near-complete insensitivity to single-nucleotide variants, with clinically relevant mutations in BRCA2 and CFTR producing AUROC scores barely above chance. The authors argue that simply importing NLP pretraining into genomics isn't sufficient, and call for biologically informed tokenisation and variant-aware training objectives before further compute is committed.

This paper introduces a nonparametric permutation test for event studies designed to remain statistically valid even when the number of firms is extremely small, addressing a key limitation of standard tests for average and cumulative abnormal returns. The method is evaluated through Monte Carlo simulations and illustrated using real-world financial data.

“How to Use the Sharpe Ratio” by Marcos López de Prado, Alexander Lipton, and Vincent Zoonekynd (ADIA Lab) re-examines the world’s most widely used measure of investment efficiency. The authors expose common statistical errors in Sharpe ratio analysis and present a rigorous framework for correcting bias, non-Normality, and multiple testing, ensuring the metric remains reliable for both researchers and practitioners.

This inaugural volume of Transactions of ADIA Lab features peer-reviewed research from ADIA Lab on finance, AI, and data science, highlighting breakthroughs in asset modelling, ethics, and high-performance computing.

This paper by Alexander Lipton & Marcos Lopez de Prado, introduces a new mathematical framework for valuing private equity (PE) investments by developing closed-form, analytical solutions to complex valuation problems, especially under conditions that more accurately reflect the reality of PE: illiquidity, nonlinearity, and market incompleteness.

By Marcos López de Prado & Vincent Zoonekynd, this paper proposes a major shift in how factor investing is approached, arguing that most traditional methods are flawed because they rely on associational rather than causal inference. By Marcos López de Prado & Vincent Zoonekynd

This paper introduces a structured, actionable framework for understanding how artificial intelligence (AI) intersects with democratic governance. Rather than treating AI as uniformly beneficial or harmful, the authors develop a dual taxonomy that maps both the risks AI poses to democracy and the positive contributions it can make.

A practical 7-step protocol to design robust, causally valid factor investing models and avoid misleading correlations in quantitative finance, by Marcos López de Prado & Vincent Zoonekynd

A roadmap for building responsible AI systems grounded in law, ethics, auditability, and governance to earn and sustain public trust in high-risk domains.

This paper explores inflation forecasting using a hedged random forest (HRF) model. Extensive empirical analysis demonstrates that this paper’s proposed approach consistently outperforms the standard random forest.

Discover how shrinkage estimators and multivariate GARCH models improve tracking-error management for benchmarked portfolios, enhancing risk control and performance

Lopez de Prado, Lipton, & Zoonekynd on factor model misspecification leading to investment inefficiency and biased portfolios. Discover why causal factor modeling is essential for sound portfolio optimization

Enhance random forest regression with optimized weighting inspired by portfolio selection. Discover a new method for improved forecasting accuracy across datasets.

Factor investing models are often misspecified, leading to biased risk premia estimates. Learn why causal inference, not associational methods, is key to accurate factor modeling.

Improving backtesting reliability in systematic investing. Resreach into walk-forward testing, resampling, and Monte Carlo simulations to avoid biases and false discoveries

Factor investing underperforms due to model specification errors, not just p-hacking. Research into how causal factor investing can improve strategy reliability and profitability.

Lopez de Prado, M., and V. Zoonekynd (2024).

ADIA Lab Research Paper Series, No. 7.

Antonov, Balasubramanian, Lopez de Prado, & A. Lipton explore a geometric approach to portfolio optimization using the Generalized Wasserstein Barycenter, offering more flexibility and rewards than Black-Litterman models.

Professor Lipton, Antonov, and M. Lopez de Prado Compare Markowitz and Hierarchical Risk Parity (HRP) portfolio allocation, & how HRP reduces noise, improves robustness, and optimizes portfolio variance.

Professor Alex Lipton explores the connection between physics and financial engineering through Kelvin waves, solving key models like Black-Scholes, Heston, and volatility swaps.

Chung, Lopez de Prado, H. Simon, and K. Wu explore feature clustering for dimensionality reduction in noisy data using parallel computing framework to optimize hyperparameters for improved model accuracy.