About me

With increasing computing power, advanced algorithms and growing data resources, machine learning methods are increasingly applied in various scientific domains. Deviating from other research fields, economic as well as financial data suffer from low signal-to-noise ratios, which significantly hinders the meaningful application of these advanced methods. This dissertation deals with the development and practical application of noise-robust algorithms in the research domain of financial economics. In particular, we focus on current problems in asset pricing, portfolio management and international finance research. The combination of both research areas (financial economics and data science) enables the discovery and practical exploitation of learnable and generalizable patterns in large, noisy data sets.

Abstract of Master Thesis (Grading 1,0)

Regarding the growing awareness of machine learning in finance, this thesis applies machine learning algorithms to equity valuation. The approach builds upon the idea that artificial intelligence tools like machine learning could contribute to valuation by offering a new, behaviorally uninfluenced perspective as opposed to financial analysts. More than 14,0001 companies were valued by several different machine learning models from January 1, 2000 to January 1, 2020 using pointin-time fundamental information. Avoiding data snooping and backtest overfitting, long-short portfolios (built by under- and overvaluation) earn returns of 24.2% per annum for the best performing random forest model compared to 12.5% for the ordinary least squares benchmark model by Bartram & Grinblatt (2018) in the time from January 1, 2000 to January 1, 2014. All models (including the Bartram and Grinblatt benchmark model) lose the signal at the latest after 2016. Using a different database with different variables and different model specifications, the loss of the signal is likely attributable to market efficiency rather than to data snooping of the chosen setup in Bartram & Grinblatt (2018).
All sophisticated Machine Learning models outperform the ordinary least squares model mainly by recognizing non-linear patterns in the data. This performance can be attributed to convergence to fundamental value rather than risk. Controlling for market betas as well as for seven other crosssectional anomalies and industry returns in time series regressions following Green et al. (2013) results in an abnormal return of 21.6% per annum for the random forest and 12% for the OLS in time series regressions from January 1, 2000 to January 1, 2014. After January 1, 2014, the signal disappears, leading to returns of -4.8% per annum for the random forest and -13.2% per annum for the OLS model from January 1, 2014 to January 1, 2020.