We use convolutional neural networks for model identification in ARMA time series models, where our networks are trained on synthetic data with known ground truths. Comparing the performance of these networks with traditional likelihood-based methods, in particular the Akaike and Bayesian Information Criteria, we are able to show that when it comes to statistical inference on ARMA orders, neural networks can significantly outperform likelihood-based methods in terms of accuracy and, by orders of magnitude, in terms of speed. We also observe improvements in terms of time series forecasting. Our approach shows the feasibility of using artificial neural networks for statistical inference in situations where classical likelihood-based methods are difficult or costly to implement.

我们使用卷积神经网络在ARMA时间序列模型中进行模型识别，在该模型中，我们的网络在具有已知地面真理的合成数据上进行训练。将这些网络的性能与传统的基于可能性的方法（尤其是Akaike和贝叶斯信息准则）进行比较，我们可以证明，就ARMA阶数的统计推断而言，神经网络在以下方面可以显着优于基于可能性的方法：精度，以及速度方面的数量级。我们还观察到时间序列预测方面的改进。我们的方法显示了在基于经典似然方法难以实施或成本高昂的情况下，使用人工神经网络进行统计推断的可行性。