발간년도 : [2023]
| 논문정보 |
|
| 논문명(한글) |
[Vol.18, No.4] DTD Transformer-MAF: Estimating the Long-Term Exchange Rates with Emphasis on Various Contextual Information at Adjacent Points in Time |
|
| 논문투고자 |
Hyuk-Jae Kwon, Minchae Kim, Ha Young Kim |
|
| 논문내용 |
The exchange rate fluctuates elastically depending on the domestic and international conditions and has a great impact on the domestic economy as well as the export and import of each country. Recently, the deterioration of Vietnam's economic situation, including the US-China conflict, has caused a major downturn in Korean exporters along with the depreciation of the Korean Won. Therefore, this study aims to help manage their exchange rate risks by predicting the trends of the exchange rates of 8 major exporting countries based on Transformer-MAF, which has shown the highest performance in multivariate exchange rate prediction studies. However, the Transformer-MAF only conveys overall contextual information to the model through an single linear embedding module, and fails to emphasize the contextual information important for predicting the fluctuation trend of each country. Therefore, we proposed a Dilated Temporal Dependency Embedding module to the Transformer-MAF to enable effective multivariate long-term prediction based on various dependency relationships at adjacent time points. The proposed embedding module consists of a high-dimensional linear projection layer that can extract comprehensive contextual information of each country, and successive randomized masking and dilated conv blocks that deliver the various contextual information from different neighboring time points in each forward propagation. Experimental results show that the proposed model has the smallest standard deviation and outperforms all comparison models with a 0.75987 CRPSsum (±0.00202), provinging that effective multivariate exchange rate forecasting is possible with only a small increase in computation. |
|
| 첨부논문 |
|
|
|
|
|
자료실 > 논문지 
