Can AI Agents Beat the Random Walk? Not So Fast

• A detectable market pattern is not necessarily an investable signal.
• Costs, regime shifts, and model constraints can erase an apparent AI edge.
• Pre-deployment tests can help managers determine whether an AI-driven strategy merits capital.

The Efficient Market Hypothesis and the random walk view have long shaped how investors think about market returns: If prices reflect available information, then systematic outperformance should be difficult, not expected.

Yet, we’ve seen the research say that some prices may contain some degree of predictability.

As AI systems become more autonomous, a familiar assumption is shaping the debate: If markets contain patterns, sufficiently advanced AI agents should eventually learn and exploit them. That assumption is attractive, but it may be incomplete. The relevant question is not simply whether markets deviate from a random walk, but whether adaptive agents can convert those deviations into positive net returns.

If markets are not fully random, adaptive agents may sometimes learn the hidden structure—but not always in an economically useful way.

Before deploying AI, investment professionals need to be able to identify a pattern that can be converted into positive net returns after transaction costs, regime shifts, noise, realistic execution constraints, and model limitations are considered.

AI may be less useful as an automatic profit engine and more useful as a diagnostic tool. It can help investors test whether a strategy is genuinely learnable, economically exploitable, and robust across market conditions.

I examine this problem in “Beyond Random Walks: Exploring the Learnability Threshold of AI Agents in Algorithmic Markets,” published in Expert Systems with Applications through the concept of a Learnability Threshold.

The idea is that there exists a point beyond which statistically detectable market patterns can no longer be profitably learned or exploited by AI agents. This threshold is especially relevant for deep reinforcement learning systems. These agents are often presented as adaptive decision-makers that can learn trading rules by interacting with market environments.

Findings show that deep reinforcement learning agents may fail to exploit long-memory market dynamics when realistic frictions are introduced. There were instances where agents converge either to inactivity or to loss-making trading behavior.

Investment professionals are familiar with this problem. A backtest may show a pattern. A factor may appear significant. A model may identify structure in returns. Yet once turnover, trading costs, slippage, market impact, regime shifts and operational constraints are included, the apparent opportunity may disappear.

AI does not eliminate this problem. In some cases, it can intensify it.

A reinforcement learning agent is trained to act. If it detects weak signals, it may try to exploit them through frequent position changes. But when those signals are close to noise, trading activity can become costly. The agent may overfit short-lived fluctuations, engage in excessive trading and erode returns through costs.

Alternatively, the agent may learn the opposite lesson. If the environment is noisy and transaction costs are meaningful, the most rational policy may be to do nothing. In this case, the AI system does not become a superior trader. It becomes inactive.

Both outcomes are important. They show that failure is not random. It is informative.

If an AI agent cannot transform a detectable pattern into positive net performance, that may tell us something about the market environment. The signal may be too weak. The cost structure may be too high. The observation space may be insufficient. The model may be too constrained. Or the pattern may be statistically real but economically untradable.

This is the essence of the Learnability Threshold.

A market can contain statistically detectable patterns without offering a tradable opportunity. This is especially true when the expected gain from exploiting the pattern is smaller than the combined cost of learning, trading, and being wrong.

This creates a practical version of what may be called a learning paradox. If a market anomaly exists but the cost of learning and exploiting it exceeds its expected return, the market may remain practically efficient for the AI agent.

For portfolio managers and quantitative researchers, this is a useful reframing. The question is not simply, “Can AI find a signal?” The better question is, “Can AI learn a policy that converts the signal into robust, positive, net-of-cost performance?”

That is a higher standard.

1. Do not be swayed simply by the power of AI. A reinforcement learning model may look more advanced than a simple heuristic, but complexity does not guarantee robustness. In noisy markets, a transparent rule-based approach can sometimes outperform a more flexible learning system because it filters noise, limits unnecessary trading, and embeds practical discipline.
2. Transaction costs are not a minor detail. They are part of the learning environment. A strategy that appears profitable in a frictionless simulation may collapse when costs are included. For AI agents, costs influence not only performance after the fact, but also the policy the agent learns. The agent may become inactive, or it may churn. Both behaviors can emerge from the same basic problem: the signal is not strong enough relative to the cost of acting on it.
3. AI should be used to stress-test strategies, not just to generate them. An agent can be placed in simulated environments with different levels of noise, persistence, cost, and regime behavior. If the strategy survives those environments, confidence may increase. If the strategy fails, the failure can reveal where the opportunity breaks down.

This makes AI valuable as a research laboratory. It can help investors ask better questions:

• Is this signal robust or fragile?
• Does it survive transaction costs?
• Does it work across regimes?
• Does the model trade because it has learned structure, or because it has overfit noise?
• At what point does a detectable pattern stop being economically learnable?

These are practical investment questions, not purely technical ones.

The question is not whether AI agents can beat the random walk in every market. They cannot. The better question is whether AI can help us understand when markets are learnable, when they are effectively random, and when the cost of learning is simply too high.