**5. Results**

Tables 1 and 2, and Figures 1–3 show adjustment levels using accuracy, the mean square error (RMSE), and the mean absolute percentage error (MAPE). In all computational methods, the level of accuracy always exceeds 82.64% for testing data, while for OLS, it reaches 75.27% for Mexico and 77.41% for Thailand. For its part, the RMSE and MAPE levels are adequate. Therefore, computational methods improve OLS by a large margin, with QNN being the one that best adjusts the result in terms of residuals (with 91.62% accuracy), followed by DNDT (with 88.10%) for Mexico. In the case of Thailand, the results improve slightly, but the order of precision is the same since the best methodology is QNN with 92.84% in test data, followed by DNDT with 89.05%. Taken together, these results provide a level of accuracy far superior to that of previous studies. Thus, in the work of [7], an accuracy of around 78.2% is revealed. In the work of [9], it is close to 73.1%, and in the study of [12], it approaches 71%. Other studies such as [1–3,5,6] achieve a precision of even less than 70%. Therefore, the difference shown by the computational methodologies applied in this study far exceeds the precision shown by the previous literature.

**Table 1.** Results of accuracy evaluation: Mexico.


**Table 2.** Results of accuracy evaluation: Thailand.



**Table 2.** *Cont.*

**Figure 1.** Results of accuracy evaluation: classification (%).

**Figure 2.** Results of accuracy evaluation: mean square error (RMSE).

**Figure 3.** Results of accuracy evaluation: mean absolute percentage error (MAPE).

These results demonstrate the greater stability offered by the QNN model compared to the rest, especially in the light of the RMSE and MAPE results obtained for three other computational methods. The results of the QNN improve the results of the popular OLS, just as it improves the precision results shown in previous works such as [9–13]. This set of computational methods observed as highly accurate represents a group of novel methods that estimate the speculative attacks and therefore different from that shown in the previous literature.

To reinforce the superiority of neural network methodologies for estimating speculative attack models, the Diebold-Mariano (DM) and Harvey-Leybourne-Newbold (HLN) tests [38,39] have been applied to compare the methodologies used and the time elapsed to perform the estimation with each of the techniques. Table 3 reports the results of the DM test, showing that all the neural network methodologies used are better options than OLS. Like QNN, it is the best option compared to the rest, since the DM test ensures that the results that exceed 1.96/ −1.96 do not reject the null hypothesis at 5% of significance, and therefore the differences observed between methodologies in the estimate are significant. On the same line, being the result with a negative sign means that the second option of the comparative is better than the second option. Likewise, the HLN test is adjusted version of DM test [39], which has better small-sample properties. Both DM and HLN tests show a significance difference between computational and statistical techniques, and the computational superiority over conventional methods. On the other hand, Figure 4 shows the average run time of the methodologies used for the estimation, where it is shown that neural network methodologies need a shorter estimation time, both for training and testing data, with QNN being the most common option efficient in terms of time use, needing 0.11 and 0.10 min to estimate with training and testing data, respectively, in the case of Mexico. For the case of Thailand, the estimate needs 0.13 and 0.11 min to estimate with training and testing data, respectively.


**Table 3.** Comparison of testing results using Diebold-Mariano (DM) and Harvey-Leybourne-Newbold (HLN) tests.

\* Indicates significance at the 5% level. \*\* Indicates significance at the 10% level.
