I benchmarked 25 models against doing nothing. None of them won (github.com)

In a groundbreaking benchmark study, 25 advanced AI and machine learning models were tested against a baseline strategy of doing nothing—specifically, holding an equal weight of stocks without any predictive capability. The results revealed a striking outcome: none of the models surpassed the simplest strategy, which itself performed dismally in a controlled environment, highlighting the challenges of achieving a net edge in financial markets when costs and unseen future data are factored in. This benchmark, termed "beat nothing," sheds light on the limitations of current AI-driven financial forecasting. The study utilized rigorous statistical methods, including a studentized circular block bootstrap, to measure true performance without bias from hindsight. The findings underscore a critical implication for the AI/ML community: that even sophisticated models struggle to outperform basic strategies when transaction costs and realistic constraints are applied. The analysis not only questions the efficacy of prevalent financial modeling approaches but also emphasizes the importance of transparent evaluation metrics, inviting a reevaluation of performance claims in algorithmic trading.