Can we predict exchange rates?

The forward premium—the difference between a forward exchange rate and the current spot rate—may contain useful information about later currency movements. The research summarised here asks whether models using that information can improve on a naïve random-walk benchmark at short horizons.

Exchange rates matter across the economy. Travellers experience them through purchasing power abroad. International investors face gains or losses when foreign assets are converted back into a home currency. Central banks monitor the domestic value of reserves and foreign-currency positions, while governments consider import prices, export competitiveness and debt obligations denominated in other currencies.

Forecasting currencies has therefore remained an important question in international finance. Yet a long line of empirical work, including influential research by Meese and Rogoff, found that models built from macroeconomic fundamentals often failed to beat a simple random walk outside the sample used to estimate them. Under that benchmark, the present exchange rate is usually the most practical starting point for predicting the next one.

Why the random walk is persuasive

Daily currency movements can be difficult to connect cleanly to economic data. Prices respond to new information, changing expectations, liquidity and risk appetite, and they sometimes move contrary to what a straightforward interest-rate comparison might imply. A random-walk model is attractive not because it explains these movements, but because more elaborate alternatives have frequently produced larger forecasting errors.

The usual empirical comparison is between this simple benchmark and predictions derived from economic theory. Historically, the benchmark often won. The research considered in this article asks whether that comparison is too narrow and whether an observed market relationship can provide a more useful competing model.

Interest parity and the forward premium

A central concept in foreign-exchange theory is uncovered interest parity. In a simplified, risk-neutral setting, the expected change in a currency should offset the interest-rate difference between two countries. If a dollar deposit paid 5% over a year and a yen deposit paid 1%, the theory would associate the four-percentage-point gap with an expected movement in the exchange rate.

Covered interest parity is an arbitrage relationship involving spot rates, forward rates and interest rates. Uncovered interest parity instead concerns the uncertain future spot rate. Classic empirical tests combine these ideas and examine whether the forward premium predicts the subsequent exchange-rate change.

The historical evidence did not generally produce the coefficient that the simplest theory predicted. In many samples, higher-interest-rate currencies appreciated rather than depreciated enough to offset their interest advantage. This recurring result became known as the forward-premium bias or forward-premium puzzle.

A different forecasting contest

If strict uncovered interest parity performs poorly, it does not necessarily follow that every part of the forward premium is uninformative. A model can instead estimate the empirical relationship found in past observations, without requiring the theoretical coefficient to equal one.

The study discussed here evaluated empirical currency models in a dynamic asset-allocation setting. Rather than considering statistical fit alone, it examined whether forecasts conditioned on the lagged forward premium could improve decisions for a risk-averse investor relative to a strategy based on the random-walk benchmark.

The comparison included models using monetary fundamentals, the forward premium and a broader information set incorporating variables such as money supply and income differences. It also allowed predicted volatility to vary over time instead of assuming that foreign-exchange volatility remained constant.

What the study found

Within the study's sample and assumptions, the forward premium supplied economically useful predictive information. Strategies that responded to both expected currency returns and changing volatility performed better on the researchers' risk-adjusted measures than a benchmark based on constant-variance random-walk innovations.

This distinction matters. Statistical predictability does not automatically survive transaction costs, estimation error, model instability or changing market conditions. Evaluating a forecast through an allocation decision provides another test: whether its information is large and dependable enough to matter to a hypothetical investor after accounting for risk.

Interpretation and limits

The evidence challenges the strongest claim that short-run exchange rates are entirely unpredictable in an economically meaningful sense. It does not establish a timeless forecasting rule. Relationships in financial markets can weaken after publication, vary by currency and period, or be offset by trading costs and risk during stressed conditions.

For policy analysis, even limited predictability could help when evaluating the sensitivity of reserves, foreign-currency debt and trade prices to exchange-rate changes. For general readers, the broader lesson is methodological: compare a model with a credible benchmark, distinguish in-sample patterns from out-of-sample performance, and consider economic value alongside statistical significance.