Forecasting with auxiliary information in forecasts using multivariate singular spectrum analysis

The internet gives us free access to a variety of published forecasts. Motivated by this increasing availability of data, we seek to determine whether there is a possibility of exploiting auxiliary information contained within a given forecast to generate a new and more accurate forecast. The proposed theoretical concept requires a multivariate model which can consider data with different series lengths as forecasts are predictions into the future. Following applications which consider published forecasts generated via unknown time series models and forecasts from univariate models, we achieve promising results whereby the proposed multivariate approach succeeds in extracting the auxiliary information in a given forecast for generating a new and more accurate forecast, along with statistically significant accuracy gains in certain cases. In addition, the impact of filtering and the use of Google Trends within the proposed methodology is also considered. Overall, we find conclusive evidence which suggests a sound opportunity to exploit the forecastability of auxiliary information contained within existing forecasts.