A1207
Title: A similarity-based approach to covariance forecasting
Authors: Mark Jennings - University of Oxford (United Kingdom) [presenting]
Chao Zhang - University of Oxford (United Kingdom)
Alvaro Cartea - University of Oxford (United Kingdom)
Mihai Cucuringu - University of Oxford (United Kingdom)
Abstract: Forecasting covariance matrices of time series is a ubiquitous problem in finance. A framework which leverages recurrent structures in data to tackle the problem is introduced. The framework calculates similarity scores between test inputs and training inputs using their recent histories and uses these scores to filter the training data. Data that do not share the dynamics of the test input are excluded from the regression, reducing the complexity of the forecasting task. Then forecasts based only on the relevant training data are produced using simple non-parametric and linear methods. Furthermore, a dynamic empirical similarity ensemble scheme is proposed, which generates a weighted average of individual forecasts based on recent performance. The framework produces computationally efficient and interpretable forecasts of the realised covariances of US equity returns and outperforms widely-used benchmark models. The framework adjusts to rapidly changing market conditions by down-weighting models that are unsuitable for the current market dynamics, which minimises the impact of model specification and leads to improved robustness against turbulent conditions compared to alternative models. The economic value of the forecasts is also evaluated by applying them to portfolio optimisation, and it is shown that the framework generates a higher Sharpe ratio than those of competing models.
