Erstellung von Prognosen zur Steuerung von Investmentstilen mit KI/ML-Methoden

Abstract

The use of artificial intelligence and machine learning techniques has continued to increase in recent years. This paper explores whether predictions about the performance of the growth and value indices of the global stock market can be made using machine learning algorithms. The growth index typically includes smaller, high-risk companies that often generate little profit but are frequently involved in new technologies and therefore have enormous growth potential. The value index, on the other hand, generally consists of well-established companies like the Bank of America, Johnson and Johnson or Procter and Gamble, which are better known for their brands like Ariel, Gillette, Pampers or Gucci. The business areas of these companies are already defined, and they are often market leaders in their respective sectors. According to prevailing economic theory, different economic environments, such as high interest rates on government bonds, inflation and other economic factors, affect these two stock groups to varying degrees. In this paper, predictions are made using machine learning algorithms such as Random Forest, Extreme Gradient Boosted Trees, and Logistic Regression to determine which of the two indices will perform better over a one-month time period. The prediction is a binary classification, the magnitude of the performance difference is not considered in the forecast. However, it was found that the results could be improved by discarding training data with minor performance differences. Because in this case, accurate predictions with a large performance difference between the two indices are more valuable than those with a small difference. The training and test data were kindly provided by an Austrian investment fund company. These data include various interest rate information, inflation data, commodity and energy prices, information on the US dollar exchange rate and the volatility index, as well as the Bitcoin price, among many other factors. Some of this data are available on a daily basis, while others only monthly. It was found that Random Forest models provided the best results, followed by XGBoost and Logistic Regression. Avoiding overfitting has been a significant challenge.

Date of Award	2024
Original language	German (Austria)
Supervisor	Ulrich Bodenhofer (Supervisor) & Rainer Haidinger (Supervisor)