Can Large Language Models Forecast Time Series of Earnings per Share? Case from Poland

This research evaluates the predictive accuracy of the cutting-edge LAG-LLaMA Large Language Model for earnings forecasts of Warsaw Stock Exchange-listed firms, comparing it with a seasonal random walk benchmark. The study uses two methods: zero-shot generalization, where the model leverages extensive pre-trained data, and fine-tuning, where historical EPS data specifically train the model. While the seasonal random walk yielded the lowest error rates, fine-tuning the LAG-LLaMA model produced comparable results in terms of MAAPE metric. The fine-tuned LAG-LLaMA model achieves the lowest RMSE and MAE errors but performs statistically equivalent to the simpler seasonal random walk model. This conclusion is specific to the Polish market and the studied period. The findings suggest LAG-LLaMA’s adaptability for longer datasets, while the simpler random walk remains effective for shorter timeframes, especially in emerging markets like Poland.