BERT-Based Prediction Model of Management Sales Forecast Error Using Japanese Firms' Earnings Meeting Transcripts

Abstract

Earnings meeting transcripts contain valuable information relevant to investment decision-making, and reflect firm's future performance such as sales revenue forecasts. In this paper, we proposed a BERT-based model to predict whether the actual sales beat the forecast using long-text Japanese EM transcripts. According to the experiment results, our proposed method outperforms the five conventional methods with an improvement ≥ 4% in accuracy.