Page 298 - Special Topic Session (STS) - Volume 3

STS544 Paolo F. et al.
results of the models which provide the lowest root mean squared error
(RMSE), the lowest mean error (ME), mean absolute error (MAE), and finally for
the model with the lowest maximum absolute error (MaxE). In addition, we
report the results for the simple forecast combination consisting of the
unweighted average of the nowcasts with mean error (in absolute terms)
below the 20th percentile. As benchmark, we report the results obtained by
using an automated ARIMA procedure.

Lowest ME Lowest RMSE Lowest MAE Lowest MaxE Combination ARIMA
ME -0.00 -0.25 -0.25 -0.25 -0.01 0.11
MAE 1.06 0.75 0.75 0.75 0.78 1.36
RMSE 1.35 0.95 0.95 0.95 0.96 1.79
MaxE 4.60 2.17 2.17 2.17 2.52 5.85
Table 1: ME, MAE, RMSE and MaxE for different nowcasting models. Lowest ME, RMSE,
MAE and MaxE indicate the models with the lowest mean error, root mean squared
error, mean absolute error and max error, respectively. The Combination column
contains performance measures for the simple nowcast combination based on the
unweighted average of our models. The set of predictors is based on firm-level
turnovers.

As we can see from Table 1, the nowcasting performance of our selected
models is better than the one of an automated ARIMA procedure. In the first
column, we report the results for the model with lowest mean error (an
automated ARIMA with principal components extracted from the firm data),
which shows a fairly poor performance in terms of MAE, RMSE and max error.
Interestingly, the same model (a boosted generalized additive model with
factors as input variables) has the best performance in terms of MAE, RMSE
and MaxE, however its mean error is fairly high (indicating biased nowcasts).
The simple combination of nowcasts shows very similar performance
compared to the best possible model in terms of MAE and RMSE, with a
slightly higher maximum error. The benefit brought by the nowcasts
combination approach is the very low mean error, which means that the
combination of nowcasts does not systematically undershoot or overshoot the
TIO. Consequently, for the rest of this paper, e.g. when we look at the results
for quarterly GDP growth, we focus on the nowcasts obtained by combining
different model predictions.
In Table 2 we report the nowcasting performance obtained when using
traffic volumes as predictors. We only report the results for the nowcasts
combination approach and for the benchmark ARIMA.

287 | I S I W S C 2 0 1 9