BEATING THE MACHINE: EVALUATING GEMINI LLM AND SEASONAL RANDOM WALK MODELS FOR EARNINGS PER SHARE FORECASTING IN POLAND
BEATING THE MACHINE: EVALUATING GEMINI LLM AND SEASONAL RANDOM WALK MODELS FOR EARNINGS PER SHARE FORECASTING IN POLAND

Summary/Abstract: Emerging markets like Poland face limited analyst coverage, with only 20% of listed companies receiving professional scrutiny, necessitating automated forecasting methods. This study investigates whether Large Language Models can outperform traditional statistical approaches in predicting earnings per share (EPS). We compare Google's Gemini LLM against the seasonal random walk (SRW) model using quarterly EPS data from 267 Polish companies (2010-2019). Performance is evaluated using Mean Arctangent Absolute Percentage Error (MAAPE), with robustness checks using RMSE and MAE metrics. Results reveal a notable discrepancy: SRW achieves superior MAAPE scores, while Gemini excels on RMSE and MAE measures. This metric-dependent performance suggests model selection should align with specific error tolerance requirements. Despite Gemini's sophisticated chain-of-thought reasoning mimicking analyst approaches, simpler models prove equally effective in markets with unsophisticated EPS dynamics. These findings challenge assumptions about LLM superiority in financial forecasting and highlight the continued relevance of parsimonious statistical methods in emerging markets.

• Details
• Contents