When the Market Changes Its Mind: Regime Fragility in Machine-Learning Forecasts of a Simulated NSE Index

Machine-learning models are widely promoted as tools for predicting stock prices, and many published studies report directional accuracies well above chance for Indian indices such as the NIFTY 50. Most of these studies test one model on one historical sample, which makes it hard to tell whether a reported edge reflects genuine skill or a favourable historical period. This paper takes a different route. I built a synthetic market calibrated to the broad statistical behaviour of a large-cap National Stock Exchange (NSE) index, in which the true amount of predictability is known and can be switched up or down. The market alternates between a calm regime, where short-term trends tend to continue, and a turbulent regime, where they tend to reverse. Four models (logistic regression, random forest, gradient boosting and a neural network) were trained with a walk-forward design on 120 independent simulated paths, each containing approximately eleven trading years and then on a further 600 paths spanning six levels of signal strength. All four models captured only 20 to 41 percent of the accuracy that was theoretically available, and every one of them was right less than half the time on turbulent days. Adding explicit regime information did not help. The simplest model, logistic regression, performed best, yet after realistic trading costs it beat a passive buy-and-hold strategy in fewer than one path in five. The results suggest that regime instability, rather than model sophistication, is the binding constraint on AI-driven market prediction.