Metaheuristics for model selection in the context of portfolio optimization

This project intends to explore the suitability of model selection techniques, inspired in those that are currently common practice in Machine Learning and Operational Research, in order to achieve better out-of-sample performance of portfolio allocation. Recent literature (DeMiguel et al. 2009, Diris et al. 2014) has highlighted the difficulty of advance portfolio allocation strategies to outperform simpler strategies, such as 1/N diversified portfolio, due to the error in the estimation of the expected returns and the variance-covariance matrix.

In econometrics, as in statistics, when a model is estimated using a multivariate volatility forecasting approach, we face the issue of trading-off flexibility of the model and accuracy. A model too flexible will have many parameters to estimate and hence the estimation will be less reliable. A model that is too simple, faces the problem of not being realistic enough to capture the features of the underlying phenomena.

In portfolio allocation problems, this trade-off still exists, but the importance of capturing the underlying phenomena can be neglected in favour of the impact in the final objective: to get a less-risky and more-profitable portfolio. Recent literature in finance and econometrics has evaluated different forecasting models in terms of the out-of-sample performance of the optimal portfolio (see for instance Engle and Colacito, 2006; Becker et al., 2015). This project will advance this research by combining ideas from three different disciplines in an innovative way. The framework for model selection will draw on ideas from Machine Learning, where the sample is usually split in three parts: training set, for estimation of parameters; validation, for model selection; and testing, for actual evaluation of the whole procedure. This framework will be adapted to the portfolio allocation problem and applied for selecting the forecasting volatility model. These volatility models will come from recent advances in Time Series and Econometrics. Finally, considering the problem from a different perspective, in order to take into account a wide range of different forecasting models, suitable Optimization techniques will be applied. The combination of these three different disciplines makes this approach unique in comparison with the state-of-the-art. For example, in a recent paper Becker et al. (2015) evaluate different loss functions in terms of their ability to discriminate the models. They follow an econometric-based approach to the problem to provide guidance for practitioners when selecting models. In contrast, our approach will propose a whole framework for selecting such models, by jointly considering the evaluation and the optimization aspects of the problem, as well as allowing the consideration of a wide variety of models. This holistic approach can be expected to bring benefits, not only in terms of better performance of the portfolio, but also a much better understanding of the problem. Note that the applicant has a wide record of publications that lays in the interface between Machine Learning and Optimization, and has recently participated in collaborations with colleagues in Finance and Econometrics. This combination of skills makes the applicant uniquely suited for pursuing this research.