Can time series foundation models predict stock prices? (github.com)

🤖 AI Summary
Recent research published on the MoneyHeap blog explores the potential of three time series foundation models to predict stock prices for 24 large U.S. stocks and ETFs. The models generated forecasts for the next-day price direction, achieving accuracy rates ranging from 49.9% to 51.1%, which is marginally below a naive prediction model that consistently suggested an upward price movement. Notably, the analysis revealed that the errors of these models were slightly larger than merely predicting no change, underscoring the challenges involved in stock price forecasting with current AI methodologies. The significance of this study lies in its assessment of model performance against benchmark predictions and random walk simulations, shedding light on the limitations of current AI models in financial markets. Interestingly, while the models showed poor correlation coefficients between their forecasts and actual stock price movements, they did demonstrate some ability to generate meaningful statistical ranges that captured real price movements approximately 72% to 80% of the time. This research contributes to the ongoing conversation around the applicability of AI and machine learning in finance, emphasizing that while these models can provide insights, they remain predominantly speculative tools rather than reliable financial predictors.
Loading comments...
loading comments...