Purpose: This study is to predict firms’ export participation based on firm heterogeneity, considering the situation where many countries around the world try to promote firms’ entry to export markets from the perspective of heterogeneous firm trade framework these days.
Research design, data, and methodology: We used the 13-year time series data from the business activity survey produced by the Statistics Korea. Total factor productivity, financial leverage, and R&D expenditure were used as input variables and export participation was used as an output variable for time series classification with deep learning. We have trained and compared the three deep learning models for time series classification: multi layer perceptron, fully convolutional network, and residual network. We implemented the models using the open source deep learning library Keras with the Tensorflow back-end. The models’ performance was evaluated using the mean of accuracy, precision, recall, and F1-score over the 10 runs on the testing data set.
Results: The results showed that the fully convolutional network (FCN) architecture performs best for the time series classification task and the recall is higher than the precision. The accuracy of the best model is 0.86, the precision is 0.64, the recall is 0.80, and the F1-score is 0.71.
Conclusions: This study contributes to promoting the understanding of deep learning approach to prediction of export participation in the context of heterogenous firm trade theory. The prediction focuses on the selection of non-exporting firms, from the perspective of policy orientation for excavating and making firms without exporting start exporting. We propose to be able to utilize the FCN for enhancing the effectiveness and efficiency of export promotion policies, in particular focused on increase in firms’ export participation, by interpreting three of the indicators being used for model evaluation, precision, recall, and F1-score, in the context of such policy.