*Postestimations*

To perform multiple-step-ahead prediction to obtain greater robustness of results, we use the iterative strategy. For this, we have trained the models for prediction for one step and two forward steps, that is, for the moments *t* + 1 and *t* +2[38]. These forecasted data for *t* + 1 and *t* + 2 are included in the data sample as actual observations. Tables 4 and 5, and Figures 5–7 point out the accuracy and residual results (RMSE and MAPE) for one-year and two-year forecasting horizons. For *t* + 1, the range of precision for the four neural networks techniques is 83.07–90.94% overall, being in the model of QNN where the percentage of accuracy is higher (90.94%) for the Mexican case. With the OLS method, the accuracy decreases to 74.72–74.90%. On the same line, for the Thai case, the precision range has been 83.34–92.63%, with QNN being again the methodology with the highest precision (92.63%). With the OLS method, the accuracy decreases to 75.64–77.15%. For *t* + 2, this

range of precision is 81.34–89.52%, being also the method of QNN in which the percentage of accuracy is higher (89.52%) for the Mexican estimations. For the OLS method, the accuracy decreases to the range of 72.78–73.81%. Moreover, in *t* + 2 for the Thai estimations, again confirms the predictive superiority of QNN (90.54%). These results show the high precision and grea<sup>t</sup> robustness of the NN techniques.

**Figure 4.** Results of time lapse for estimation.

**Table 4.** Multiple-step ahead forecasts in forecast horizon = *t* + 1 and *t* + 2 (Mexico).



**Table 5.** Multiple-step ahead forecasts in forecast horizon = *t* + 1 and *t* + 2 (Thailand).

**Figure 5.** Multiple-step ahead forecasts in forecast horizon: accuracy.

**Figure 6.** Multiple-step ahead forecasts in forecast horizon: RMSE.

**Figure 7.** Multiple-step ahead forecasts in forecast horizon: MAPE.
