Machine Learning Shows Fundamentals Do Predict Exchange Rates

Image: "Currency Exchange Corporation, Oakwood tube station" by Philafrenzy (CC BY-SA 4.0) via Wikimedia Commons, recoloured.

Do economic fundamentals help forecast short-run exchange rate moves? A new study in the Journal of Financial and Quantitative Analysis says yes, once you let their effects vary with global financial conditions. Using machine learning on high-dimensional country data, the authors find fundamentals matter most when currency market volatility is high.

Letting the data decide when fundamentals matter

For decades, international finance has lived with an awkward result: models built on economic fundamentals rarely beat a simple random walk when forecasting exchange rates a short time ahead and out of sample. This exchange rate predictability puzzle has survived many attempts to resolve it, and it has encouraged the view that currency movements in the short run are mostly noise relative to the macro variables that theory says should drive them.

Ilias Filippou, David E. Rapach, Mark Taylor and Guofu Zhou approach the puzzle as a measurement problem. Instead of assuming that each fundamental has a fixed effect, they interact a rich set of country characteristics with global state variables, allowing the relevance of each characteristic to change with the wider financial environment. Because this creates a very large number of candidate predictors, they rely on machine-learning techniques to handle the high-dimensional data and to produce out-of-sample forecasts.

According to Filippou, Rapach, Taylor and Zhou (Journal of Financial and Quantitative Analysis, 2026), fundamentals do matter for short-run exchange rates, but their effects depend on global financial conditions. The interactions between country characteristics and foreign exchange market volatility turn out to be particularly important. The authors interpret this pattern as a "flight to fundamentals": when financial stress rises, investors appear to pay closer attention to the economic characteristics of the countries whose currencies they hold.

The authors also report that their forecasts deliver substantial economic value, rather than only statistical gains, and that the results offer new insight into the carry trade, the strategy of borrowing in low-interest currencies to invest in high-interest ones. The abstract does not give the specific magnitudes, sample of currencies or time period, so readers who want the size of the gains or the exact set of models should consult the full article.

As with any forecasting study, the evidence is tied to the countries, variables and period the authors examine, and machine-learning predictions identify which interactions forecast well rather than proving the underlying causal channel. The flight-to-fundamentals reading is the authors' interpretation of the predictive pattern.

Why it matters: For researchers in finance, economics and management, the study is a clear example of machine learning used to answer a long-standing theoretical question rather than only to improve accuracy. Its lesson travels beyond currencies: when a variable seems useless on average, the problem may be that its effect is conditional, and flexible high-dimensional methods can reveal when it matters. The design, interacting unit-level characteristics with aggregate state variables, is a template that can be applied to other prediction puzzles in the social sciences.

---

Filippou, I., Rapach, D. E., Taylor, M., & Zhou, G. (2026). Economic Fundamentals and Short-Run Exchange Rate Prediction: A Machine-Learning Perspective. Journal of Financial and Quantitative Analysis. Read the original paper. Independent editorial summary of third-party research.