| 總結: | <p>This thesis considers the question of whether machine learning models can utilise the data contained in past financial statements to predict the market reaction to a future earnings announcement. The theoretical motivation for this hypothesis is drawn from a literature review encompassing the research domains of financial statement analysis, and earnings forecasting. </p>
<p>For the empiric evaluation of the hypothesis, a range of machine learning models and traditional linear models are trained to predict the abnormal return of a stock following its earnings call based on 121 variables from the balance sheet, income-, and cashflow-statement. These quarterly financial statement variables for the entire North American stock market are taken from the Compustat FUNDQ data set between the years 1991 and 2017, and the critical problem of missing values in the data is addressed with a comprehensive pre-processing strategy. The abnormal stock return serving as the dependent variable is computed as the Buy-And-Hold Abnormal Return (BHAR) of a particular stock until 30 days after the earnings announcement. </p>
<p>The design of the experiments involves model setups formulated as regressions and classifications, which are evaluated using an out-of-time and out-of-sample test sample. For it, the models demonstrate an ability to predict the correct sign of the abnormal market reaction in the majority of the cases using a newly introduced metric called PC. Additionally, the model predictions are demonstrated to translate into positive quarterly returns when employed in a simulated trading strategy over the study period. These task-specific evaluation metrics confirm a superior performance of the random forest and neural network over traditional linear models like the Lasso regression.</p>
|
|---|