Following interactive plot allows to compare daily prediction errors of main studied models.

​

The plot represents mean absolute percentage error (MAPE) of studied models. Each color identifies a bimestral, while the height is directly proportional to the prediction error itself.

​

 

Errors are split in six groups of two months each, and can be filtered through the first filter in top-right corner.

It is possible to use the second filter to activate the de-seasonalization procedure described here.

​

 

One can notice that the Artificial Neural Network (neural) prediction is always better compared to the other ones just if the de-seasonalization is active (DESEASON = LogAdditive). This method stands as a "standardization" technique for output variables, allowing the net to adapt quickly.

​

Gaussian Process Regression (gaussian) is able to obtain a good MAPE compared to other models mainly using an Exponential kernel (Exp). Compared to the Neural Net, it has a higher prediction error, but a more reliable prediction confidence interval structure.

​

 

Linear model (linear) and Nearest Neighbour regression (nearest) show a good prediction power with respect to the MAPE, preferring fewer parameters and easier parameters explanation instead of a higher forecasting precision for Neural Nets and reliable confidence intervals for Gaussian Processes.