Beyond Random Walks: Exploring the Learnability Threshold of AI Agents in Algorithmic Markets

Financial markets are often assumed to be random and unpredictable, yet their structure increasingly reflects the influence of automated decision systems. This study examines the practical limits of machine learning in such environments by introducing the concept of a learnability threshold a boundary where computational agents can no longer extract meaningful patterns from market data despite apparent statistical regularities. Using a comparative simulation framework, the paper evaluates two distinct approaches to decision-making: a transparent, rule-based trading heuristic and a learning agent trained to replicate it. The analysis spans both structured and noisy market regimes, allowing for a systematic examination of how market complexity and transaction frictions affect performance. Results show that when the signal-to-noise ratio of market information falls below the learnability threshold, adaptive models cease to improve outcomes, while simple heuristic rules continue to perform consistently. The findings suggest that algorithmic systems, rather than replacing human expertise, can serve as diagnostic tools for identifying when learning is feasible and when judgment-based reasoning remains superior, offering a balanced perspective on the evolving relationship between computational decision systems and the adaptive logic of real markets. JEL Codes: C45; C63; D83; G14; G17