Nowcasting in Tunisia using large datasets and mixed frequency models

The object of this paper is to nowcast, forecast and track changes in Tunisian economic activity during normal and crisis times. The main target variable is quarterly real GDP (RGDP) and we have collected a large and varied set of monthly indicators as predictors. We use several mixed frequency models, such as unrestricted autoregressive MIDAS (UMIDAS-AR), three pass regression filter (3PRF) and mixed dynamic factor models (MDFM). We evaluate these models by comparing them with benchmarking low frequency models including vector autoregressive (VAR) and ARMA models. The dynamic factor and the 3PRF forecasts are more accurate in terms of mean squared errors (MSE) than other alternatives models both in-sample and out of sample in normal times, meaning before the COVID19 period. Forecast errors derived from low frequency models including crisis periods are larger than errors from mixed data sampling approaches including autoregressive terms due mainly to the failure of the low frequency models to capture these tail events. Fortunately, the reliability of nowcasts and forecasts increase when using the mixed frequency dynamic factor model based on information at both monthly and quarterly frequencies.