Beating the Machine: Why Simple Random Walks Outshine Google's Gemini in Forecasting Earnings on Poland's Exchange.

This study investigates the application of Large Language Models (LLMs) for forecasting Earnings Per Share (EPS) on the Warsaw Stock Exchange, comparing Google's Gemini model against the seasonal random walk (SRW) approach. Analyzing quarterly EPS data from 267 Polish companies (2010-2019), the mean arctangent absolute percentage error (MAAPE) is employed to address limitations of traditional error metrics. Results demonstrate that despite Gemini's sophisticated capabilities, the simpler SRW model consistently produces lower error rates when measured by MAAPE, although Gemini outperforms on RMSE and MAE metrics. The divergence highlights how model selection should be guided by specific error tolerance requirements. Additionally, we explore Gemini's chain-of-thought reasoning and certainty scores to assess model confidence. These findings contribute to the understanding of EPS prediction in emerging markets with limited analyst coverage and suggest that complex models may not provide significant advantages in markets characterized by relatively unsophisticated EPS dynamics.