| Abstract [eng] |
The objective of the master’s thesis is to create a machine learning model that predicts whether a stock price will exhibit bullish or bearish tendencies after the Quarterly Earnings Call release. With a tailored trading strategy, the model aims to generate a positive return, ideally surpassing that of a passive index buying strategy. To achieve this, a dataset containing 33,362 earning call transcripts from 685 companies span ning the period from January 1, 2005, to March 31, 2025, was used for transcript and speaker sen timent extraction using the pretrained BERT, FinBERT, and GPT5nano models. Sentiment data was enriched with stockrelated data from the Yahoo Finance API before the Earnings Call to capture the quarter’s price movement tendencies. Further data analysis, feature engineering, and data set clean ing were performed. To obtain robust results, a few model categories were tested using the same strategy: Lasso and Ridge regression, XGBoost, LightGBM, and Random Forest to identify which can better predict price movement. The results indicate that earnings calls can serve as shortterm catalyst events, and combining sentiment with financial metrics provides an additional predictive signal. While classification perfor mance is low due to market noise, small predictive advantage and sector awareness yielded positive returns at times, beating the market index. Results revealed that increased model complexity does not improve performance, as regularised linear models, such as Ridge regression, delivered more robust results. |
