Exchange rates forecasting and trend analysis after the COVID-19 outbreak: new evidence from interpretable machine learning

ABSTRACT

We investigate the predictability of 12 exchange rates with machine learning, Deep Learning and interpretable machine learning (IML) models, based on a daily dataset from December 2019 to August 2021. We find that the appreciation and depreciation of exchange rates can be partly captured by Light Gradient Boosting Machine (LightGBM) and Long Short-Term Memory, especially for the developed currencies. Inconsistent with general perception, the LightGBM model performs the best in exchange rates forecasting since its short-term information extracting mode and great robustness on small datasets. Furthermore, by employing a representative global IML method, the Accumulated Local Effect algorithm, we find that the 1 ~ 3 lags of exchange rates provide more useful information for forecasting, which can help investors improve their models’ predictive ability.

KEYWORDS:

• Exchange rates
• forecasting
• machine learning
• interpretable machine learning

JEL CLASSIFICATION:

• C53
• F37

Acknowledgments

This research is financially supported by the National Natural Science Foundation of China under projects No. 72173145 and No. 72134002, and the Beijing Technology and Business University under projects No. QNJJ2020-36 and No. PXM2020_014213_000017.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Notes

¹ https://scikit-learn.org/stable/modules/generated/sklearn.model_selection.TimeSeriesSplit.html.

Additional information

Funding

This work was supported by the Beijing Technology and Business University [No. QNJJ2020-36 and No. PXM2020_014213_000017]; National Natural Science Foundation of China [No. 72173145 and No. 72134002].