Estimation of risk premia from option data and using machine learning methods: comparison, forecast quality and potential of hybrid strategies

点击翻译按钮获取中文摘要

The explanation of risk premia, in terms of their time series properties and in the cross-section of traded assets, is at the heart of financial economics. While the fundamental asset pricing equation makes a clear statement about the economics of risk compensation - it is the covariance of an asset's return and the stochastic discount factor that determines the risk premium - empirical implementations of this general concept are challenging and continue to spur theoretical and econometric research in finance. Within a vast and active literature one can distinguish two strategies. The first employs theory-based structural models for their empirical analysis, which has the advantage of being based on principled economic thought. However, the assessment of the empirical performance is often hampered by intricate model structures that preclude the use of standard econometric methods. Moreover, the models employed are sometimes highly stylized and rely on apparently unrealistic assumptions. The second strategy consists of empirical approaches that are econometrically more accessible, but are prone to the critique of measurement without theory and an undisciplined fishing for risk factors. Joining the forces of two finance research groups at the Universities of Frankfurt and Tübingen, this project takes a closer look at two novel frameworks to measure risk premia that can be conceived of as extreme cases of the theory-based and the empirical strategies. The first is forward-looking, because it exploits market expectations that are reflected in option prices. It is theory-based, because it relies on a reformulation of the fundamental asset pricing equation. The non-parametric nature of this strategy counters the critique of employing unrealistic assumptions. The second approach employs machine learning methods and is thus backward-looking in a sense that these methods look for patterns in historical data. One does not draw on economic theory, but concepts from data science. While their philosophies are fundamentally different, the two new frameworks are concerned with the same object of interest, namely the risk premium reflected in the conditional expected return of a financial asset. This common objective makes the two methodologically disjoint approaches comparable, and in principle combinable. Accordingly, our proposal aims at providing a comparative evaluation of the two frameworks in terms of their forecast performance - recalling that the conditional expectation is the mean-squared-error optimal forecast - and the development and assessment of hybrid frameworks that combine the financial theory- and data science-based approaches. Because a lack of deeper understanding of the limits of quantitative models was one main driver of the recent financial crises, we emphasize the need to provide a critical view on the possibilities and limitations of the option-based and the data science-based framework, as well as the hybrid models to be developed.