A novel (U)MIDAS-SVR model with multi-source market sentiment for forecasting stock returns

From the point view of behavioral finance, market sentiment plays an important role in forecasting stock returns. How to accurately measure the impact of market sentiment is a challenge work. Two issues on nonlinear relationship and mixed-frequency data have to be addressed. To this end, we introduce methods of mixed-frequency data into SVRs and develop a novel (U)MIDAS-SVR model. It can be estimated by solving the Lagrange duality technique of quadratic programming. We then apply the (U)MIDAS-SVR model to predict weekly returns of SHSE and SZSE in China using the mixed-frequency market sentiment as covariates. The empirical results show that the (U)MIDAS-SVR model is promising and MIDAS-SVR is superior to those competing models in terms of MAE and RMSE. In addition, we design seven scenarios by considering different data source combinations and find that the multi-source market sentiment is helpful to improve forecasting performance on stock returns.