We propose EcoTrans, an open-source framework for macroeconomic forecasting with interpretability, using a Transformer-based interface to support massive multimodal information access.
The data file is the example data, and you can run it directly by example.py.You can choose the type of external data you want, and there is no limit to the amount of external data you can add. Datasets are composed of three column fields, Unique_ID indicates the type of data, ds means the time,y for data to be predicted, and ex1 to ex N for external data. Refer to the readme in the Data folder for more detailed descriptions.
It should be noted that what is given in the example.py file is the direct code of the paper's experiment, where circular cutting is adopted to cut the test set. You may not need to do this in a real application scenario. You can remove some configuration loops from the example.py file, which we present in our new example in example2.py.
You can watch the performance of the model change by adjusting the relevant parameters in example.py and config_function_ex.py. It should be noted that our model supports long-period prediction, but in our experiment, the prediction time is 1, which means that the results of all test sets need to be implemented through a loop. See the comments in example2.py for details.
EcoTrans has proposed an extensible external data access method. After testing, with the increase of external data access quantity, our approach has stronger robustness than the baseline method, and the model also has a certain degree of interpretability.
We compared the performance of our model in different countries and across different types of data, and the overall result was more accurate than the baseline model. Our model can be well-fitted to the external data information, and we have shown in the sample how to get the decomposed results, which you can do directly.
table1 Ablation study over EcoTrans model’s performance on different countries
| China | United States | United Kindom | Japan | |
|---|---|---|---|---|
| N-BEATS (with no external data) | 0.008103 | 0.004641 | 0.006511 | 0.004799 |
| EcoTrans (with no external data) | 0.006823 | 0.033024 | 0.006764 | 0.004787 |
| N-BEATSx (5 series) | 0.059242 | 0.015212 | 0.007450 | 0.006196 |
| EcoTrans (5 series) | 0.008331 | 0.004304 | 0.006784 | 0.004936 |
| N-BEATSx (10 series) | 0.058364 | 0.011632 | 0.007810 | 0.004550 |
| EcoTrans (10 series) | 0.007927 | 0.004418 | 0.006696 | 0.004770 |
| N-BEATSx (15 series) | 0.044965 | 0.283166 | 0.007845 | 0.051633 |
| EcoTrans (15 series) | 0.008460 | 0.004237 | 0.006805 | 0.004810 |
| N-BEATSx (20 series) | 0.050844 | 0.228063 | 0.008979 | 0.022131 |
| EcoTrans (20 series) | 0.008108 | 0.004371 | 0.006809 | 0.005188 |
Table2 The generalizability of EcoTrans for forecasting various macroeconomic indicators
| PMI | Imports | Exports | GR | PGO | CI | IE | Tax | |
|---|---|---|---|---|---|---|---|---|
| BVAR | 0.0039 | 0.0161 | 0.0223 | 0.0352 | 0.0182 | 0.0324 | 0.0189 | 0.0361 |
| N-BEATS | 0.0053 | 0.0235 | 0.0169 | 0.0651 | 0.0077 | 0.0128 | 0.0238 | 0.0565 |
| N-BEATSx | 0.1594 | 0.0721 | 0.0691 | 0.0488 | 0.0074 | 0.0105 | 0.0232 | 0.0591 |
| EcoTrans | 0.0030 | 0.0108 | 0.0222 | 0.0339 | 0.0065 | 0.0108 | 0.0208 | 0.0420 |
The numbers in the table represent the SMAPE error
you can directly use example.py