Can AI predict EUR/USD? What forty years of research say

The promise comes up regularly in sales demos: an artificial intelligence model, fed with market data and news, would tell you where EUR/USD will be in three or six months. For a finance department the stakes are real — a good call would let it hedge at the best moment.

The question has been studied for four decades, long before generative AI. The results are stable enough to answer clearly: the direction of the rate cannot be forecast in a usable way over a corporate hedging horizon; its size can.

The starting point: the forecast "the rate will not move"

In 1983, Richard Meese and Kenneth Rogoff compared the exchange rate models of the time — based on money, rates, growth — with a naive rule: the best forecast of tomorrow's rate is today's rate. Outside the sample used to build them, none of those models did better than this "random walk".

Thirty years later, Barbara Rossi's reference review took stock of everything tried since. Her conclusion fits in two words: "it depends". Some predictors, at some horizons, over some periods, improve forecasts slightly; the random walk remains the hardest benchmark to beat, and the gains do not always survive a change of period or method.

Professional forecasters do no better

If models fail, do market economists succeed? Studies comparing their published forecasts with the rates actually observed say, for the most part, no.

• On the Wall Street Journal's panel of economists, most forecasters do not predict exchange rates better than a random walk (Mitchell and Pearce, 2007).
• In a Swedish survey of market participants, none of the regular respondents significantly beats the random walk on the krona against the euro and the dollar (Kiss et al., 2023).

What about machine learning?

It is the legitimate question: do more powerful methods find what classical models missed? A study published in 2018 shows that simple machine-learning methods applied to economic fundamentals beat the no-change forecast for major currencies between 1973 and 2014 — at a one-month horizon (Amat, Michalski and Stoltz).

The result is serious, and must be read for what it is: a gain measured at one month, on historical data, on average over a long period. It does not say that a model can call EUR/USD six or twelve months out, the horizon of a hedging policy, nor that an average gain translates into the right decision at the moment the company has to hedge.

As for chat assistants, they are language models: they are built to produce plausible text, not to forecast a financial series. Asked about the future rate, they return a number that looks like a forecast, without the data or method that would let you judge its quality.

The forward rate is not a forecast either

A common confusion is to read the forward rate as the market's expectation. It is not: the forward rate is derived from the spot rate and the interest-rate differential between the two currencies, simply because there must be no arbitrage. This is known as covered interest parity.

Studies confirm it forecasts poorly: as early as 1984, Eugene Fama showed that the gap between forward and spot does not correctly predict the subsequent move, and often points the opposite way. This is the "forward premium puzzle", still debated today.

What can be forecast: the size of moves

While direction escapes models, the size of moves is forecast far better. The review by Ser-Huang Poon and Clive Granger, comparing 93 studies, concludes that implied volatility taken from option prices forecasts future volatility better than statistical models based on history.

This is directly usable. Under the usual normal-distribution assumption, an annual volatility converts into a range at a given horizon by multiplying it by the square root of the time in years.

This range does not say where the rate will go; it says how far it can reasonably go. That is exactly what a finance director needs to set a budget rate, size a hedge ratio or price an adverse scenario.

Questions to ask a tool that claims to forecast

What this changes for a finance department

1. 1
   Think in ranges

   Replace the question "where will the rate be?" with "how far can it go, and what does that cost my margin?". Implied volatility gives the answer, maturity by maturity.

2. 2
   Set a budget rate that allows for the range

   A budget rate equal to today's rate will be broken one time in two. Setting it with the range in mind makes the target achievable.

3. 3
   Hedge by rule, not by conviction

   A hedge ratio by horizon, decided in advance, avoids postponing a hedge on the strength of a forecast — human or algorithmic.

4. 4
   Use AI where it helps

   Reading documents, consolidating exposure, preparing committees: there it saves time without claiming what it cannot do.

Sources

• Meese, R. and Rogoff, K. (1983), "Empirical exchange rate models of the seventies: Do they fit out of sample?", Journal of International Economics, 14, pp. 3-24.
• Rossi, B. (2013), "Exchange Rate Predictability", Journal of Economic Literature, 51(4), pp. 1063-1119.
• Mitchell, K. and Pearce, D. (2007), "Professional forecasts of interest rates and exchange rates: Evidence from the Wall Street Journal's panel of economists", Journal of Macroeconomics, 29(4), pp. 840-854.
• Kiss, T., Kladivko, K., Silfverberg, O. and Österholm, P. (2023), "Market Participants or the Random Walk – Who Forecasts Better? Evidence from Micro Level Survey Data", Working Paper 2023:2, Örebro University.
• Amat, C., Michalski, T. and Stoltz, G. (2018), "Fundamentals and exchange rate forecastability with simple machine learning methods", Journal of International Money and Finance, 88, pp. 1-24.
• Fama, E. (1984), "Forward and spot exchange rates", Journal of Monetary Economics, 14(3), pp. 319-338.
• Borio, C., McCauley, R., McGuire, P. and Sushko, V. (2016), "Covered interest parity lost: understanding the cross-currency basis", BIS Quarterly Review, September 2016.
• Poon, S.-H. and Granger, C. (2003), "Forecasting Volatility in Financial Markets: A Review", Journal of Economic Literature, 41(2), pp. 478-539.