Can AI Predict the Stock Market?

“Can AI predict the stock market?” is too broad to test. Specify the asset universe, forecast target, horizon, information available at decision time, and benchmark. Predicting volatility, classifying a market regime, estimating execution cost, and predicting short-horizon return are different tasks. A result for one target does not establish skill on another, and a statistically accurate forecast may still be unprofitable after fees, spreads, slippage, capacity limits, and risk. The efficient-markets literature is a framework about how prices reflect information, not a statement that every market is perfectly efficient. Fama’s review discusses the theory and empirical tests. If a public signal becomes useful, other participants may trade on it and change its value. That feedback can make a signal decay, a process often called alpha decay. It does not prove that no forecast can ever work; it means claims need a defined market, period, information set, and realistic test. Repeatedly testing many features and parameter choices raises the chance that the best historical result reflects chance rather than durable signal. Bailey and coauthors describe this selection problem in their Probability of Backtest Overfitting paper. A credible evaluation keeps timestamps point-in-time, chooses baselines before examining final results, preserves later periods for evaluation, includes realistic costs, and discloses how many variants were tested. A single strong backtest is not evidence of guaranteed returns. The SEC, NASAA, and FINRA investor alert warns that purported AI trading systems are used in pitches promising high or guaranteed returns. Verify claims and registration rather than relying on model branding.