Hybrid FinBERT-LSTM Deep Learning Framework for Stock Price Prediction: a Sentiment Analysis Approach Using Earnings Call Transcripts

EasyChair Preprint 15694

Abstract

Stock market prediction remains a critical area of research due to its significant economic implications and inherent complexity. With advancements in machine learning, research interest has grown substantially in understanding the impact of textual data on financial forecasting. This study presents a hybrid FinBERT-LSTM model that combines sentiment analysis of quarterly earnings conference calls with traditional price prediction methods. We evaluate our model’s effectiveness against standalone LSTM approaches across six major US stocks from the financial and technology sectors. Experimental results demonstrate that the sentiment-enhanced hybrid model achieves superior predictive accuracy for four of the six studied stocks, as measured by Mean Squared Error (MSE), Mean Absolute Percentage Error (MAPE), and Accuracy metrics. Most notably, Citibank and Meta demonstrated substantial improvements when incorporating sentiment analysis, with MSE scores approximately 38 percent lower compared to predictions without sentiment data. Our findings contribute to the growing body of research on textual analysis in financial forecasting and offer practical implications for investment decision-making.

Keyphrases: Earnings Conference Calls, FinBERT, Sentiment Analysis, Stock price prediction, deep learning

@booklet{EasyChair:15694,
  author    = {Biswadeep Sarkar and Abdul Shahid},
  title     = {Hybrid FinBERT-LSTM Deep Learning Framework for Stock Price Prediction: a Sentiment Analysis Approach Using Earnings Call Transcripts},
  howpublished = {EasyChair Preprint 15694},
  year      = {EasyChair, 2025}}