Forecasting investment: A fishing contest using survey data

This paper assesses the usefulness of business surveys as a source of information for investment developments in Portugal. This will be achieved by what will be named a “fishing contest”, where the “participants” are bridge models, models based on principal components (derived from standard and non-standard methods), and models built with the outcome of partial least squares regressions. All models, based on quarterly data, are estimated using a general-to-specific approach and are designed to produce 1 to 4 out-of-sample direct forecasts. The accuracy of these forecasts is then compared with the one of autoregressive processes. The empirical evidence indicates that, in general, there is always a participant in the fishing context that produces a lower out-of-sample Root Mean Squared Error (RMSE) than the one associated with the autoregressive benchmark. In most cases, the combination of autoregressive processes with each participant reduces the RMSE further. A striking outcome is the relative accuracy of bridge models.