Tizian Otto

This paper compares bootstrap simulation approaches in the context of the maximum drawdown (MDD) risk measure for stock market and cryptocurrency returns. Our comparisons are based on the complete distribution of the MDD using stochastic dominance tests. The standard Efron (1979) bootstrap severely underestimates the true MDD. The simulation results of the moving block bootstrap approach are reasonably good as long as the stationarity problem does not become striking. The stationary bootstrap approach of Politis and Romano (1994) provides the best results. Investment practitioners should choose the Politis and Romano (1994) method as their first choice to model MDD risk.

This paper evaluates the predictive performance of machine learning methods in forecasting European stock returns. Compared to a linear benchmark model, interactions and nonlinear effects help improve the predictive performance. But machine learning models must be adequately trained and tuned to overcome the high dimensionality problem and to avoid overfitting. Across all machine learning methods, the most important predictors are based on price trends and fundamental signals from valuation ratios. However, the models exhibit substantial variation in statistical predictive performance that translate into pronounced differences in economic profitability. The return and risk measures of long-only trading strategies indicate that machine learning models produce sizeable gains relative to our benchmark. Neural networks perform best, also after accounting for transaction costs. A classification-based portfolio formation, utilizing a support vector machine that avoids estimating stock-level expected returns, performs even better than the neural network architecture.

Machine learning-based market beta estimators outperform established benchmark models both statistically and economically. Analyzing the predictability of time-varying market betas of U.S. stocks, we document that machine learning-based estimators produce the lowest forecast and hedging errors. They also help create better market-neutral anomaly strategies and minimum variance portfolios. Among the various techniques, random foests perform best overall. Model complexity is highly time-varying. Historical stock market betas, turnover, and size are the most important predictors. Compared to linear regressions, allowing for nonlinearity and interactions significantly improves predictive performance.

This paper uses a comprehensive set of predictor variables from the five largest Eurozone countries to compare the performance of simple univariate and machine learning-based multivariate models in forecasting stock market crashes. In terms of statistical predictive performance, a support vector machine-based crash prediction model outperforms a random classifier and is superior to the average univariate benchmark as well as a multivariate logistic regression model. Incorporating nonlinear and interactive effects is both imperative and foundation for the outperformance of support vector machines. Their ability to forecast stock market crashes out-of-sample translates into substantial value-added to active investors. From a policy perspective, the use of machine learning-based crash prediction models can help activate macroprudential tools in time.

This paper examines the cross-sectional properties of stock return forecasts based on Fama–MacBeth regressions using all firms contained in the STOXX Europe 600 index during the September 1999–December 2018 period. Our estimation approach is strictly out of sample, mimicking an investor who exploits both historical and real-time information on multiple firm characteristics to predict returns. The models capture a substantial amount of the cross-sectional variation in true expected returns and generate predictive slopes close to one, i.e., the forecast dispersion mostly reflects cross-sectional variation in true expected returns. The return predictions translate into high value added for investors. For an active trading strategy, we find strong market outperformance net of transaction costs based on a variety of performance measures.

This paper tests the predictive performance of machine learning methods in estimating the illiquidity of U.S. corporate bonds. Machine learning techniques outperform the historical illiquidity-based approach, the most commonly applied benchmark in practice, from both a statistical and an economic perspective. Tree-based models and neural networks outperform linear regressions, which incorporate the same set of covariates. Gradient boosted regression trees perform particularly well. Historical illiquidity is the most important single predictor variable, but several fundametal and return- as well as risk-based covariates also possess predictive power. Capturing interactions and nonlinear effects among these predictors further enhances predictive performance.