Earnings prediction using machine learning methods and analyst comparison

• Autor(a) principal: Martins, Alexandre Inês
• Data de Publicação: 2022
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: http://hdl.handle.net/10400.14/38654
• topic: Earnings announcements; Analyst errors; Event study; Machine learning; Technical analysis; Anúncio de resultados; Erros dos analistas; Estudos de eventos; Análise técnica; Domínio/Área Científica::Ciências Médicas::Outras Ciências Médicas

In the course of this dissertation we propose an experimental study on how technical, macroeconomic, and financial variables, alongside analysts’ forecasts, can be used to optimize the prediction for the subsequent quarter’s earnings results using machine learning, comparing the performance of the models to analysts’ forecasts. The dissertation includes three steps. In step one, an event study is conducted to test abnormal returns in firms’ stock prices in the day following earnings announcement, grouped by earnings per share (EPS) growth in classes of size 3, 6 and 9, computed for each quarter. In step two, several machine learning models are built to maximize the accuracy of EPS predictions. In the last step, investment strategies are constructed to take advantage of investors’ expectations, which are closely correlated with analysts’ predictions. In the backdrop of an exhaustive analysis on quarterly earnings predictions using machine learning methods, conclusions are drawn related to the superiority of the CatBoost classifier. All machine learning models tested underperform analyst predictions, which could be explained by the time and privileged information at analysts’ disposal, as well as their selection of firms to cover. Regardless, machine learning models can be used as a confirmation for analyst predictions, and statistically significant investment strategies are pursued with those fundamentals. Importantly, high confidence predictions by machine learning models are significantly more accurate than the average accuracy of forecasts.