Adaptive Markets Hypothesis and Conditional Ecology Performance: Evidence from Machine-Learning Market-Timing Strategies

Abstract

We evaluate traditional and modern machine-learning market-timing strategies from 1970 to 2024 under a common information set, portfolio rule, and realistic transaction costs. Unconditionally, none of the strategies delivers robust abnormal performance relative to buy-andhold, consistent with the Efficient Markets Hypothesis (EMH). Conditioning on an ex-ante ecology vector capturing crowding, funding, liquidity, and volatility, performance, risk, and cross-strategy co-movement vary systematically, with different model families specializing in distinct habitats. These state-dependent patterns are difficult to reconcile with a purely risk-based EMH but align closely with the Adaptive Markets Hypothesis (AMH).