The predictive power of earnings conference calls : predicting stock price movement with earnings call transcripts
MetadataShow full item record
- Master Thesis 
Earnings conference calls are considered a valuable text based information source for investors. This paper investigates the possibility to predict the direction of stock prices by analyzing the transcripts of earnings conference calls. The paper investigates 29 339 different earnings call transcript from 2014 to 2017 and classify the individual documents to either be part of class up or down. Four different machine learning algorithms are used to classify and predict based on the bag of words method. These machine learning algorithms are Naive Bayes, Logistic regression with lasso regularization, Stochastic Gradient Boosting, and Support Vector Machine. All models are compared to a benchmarks based on S&P 500. The model with best performance is logistic regression with a classification error of 43,8%. In total, 2 of 4 models beats the benchmark significantly, namely logistic regression and gradient boosting. With these results, the paper concludes that earnings calls contain predicting power for next day’s stock price direction.