What Changed for Exchange Rate Models, and What Did Not

In 1983, Richard Meese and Kenneth Rogoff published a paper that embarrassed an entire field. They compared structural exchange rate models — models built on interest rates, money supply, income differentials — against a simple random walk: the prediction that tomorrow's exchange rate will equal today's. The structural models lost. At every horizon, the random walk was as good or better.

This result is the Meese-Rogoff puzzle, and it has never been cleanly overturned. FX models had no forecasting edge. The implication was uncomfortable then and has not fully gone away: if exchange rates cannot be predicted out of sample by macroeconomic fundamentals, what exactly is macro-driven FX analysis doing?

What Changed

What changed is not that economists built better models, and not — as this piece once claimed — that the puzzle went away. What changed is how well the models describe the data, and the reason is that the world changed: specifically, central banks did.

The 1980s and early 1990s were characterized by inconsistent inflation regimes. Central banks in most developed countries were still establishing credibility, inflation targets were either absent or routinely breached, and the relationship between monetary policy and the price level was noisy. In that environment, rate differentials were weak signals: you could not trust that a rate gap would persist, or that inflation would behave in the way the models assumed.

The credible inflation targeting era that followed — adopted by New Zealand in 1990, the United Kingdom in 1992, Sweden and Canada in 1993, and gradually by most advanced economies over the next decade — changed the regime. When a central bank commits credibly to a 2% target and actually hits it over multiple cycles, the relationship between policy rates and inflation becomes predictable. And when that relationship is predictable, the models built on rate differentials describe currency movements far better than they did before.

What Engel and Wu Actually Found

Charles Engel and Steve P. Y. Wu (2024) put numbers on this. A standard model — real interest rates and expected inflation for both countries, the US comprehensive trade balance, and measures of global risk and liquidity demand — is, in their words, well supported in the data for the dollar against the other G10 currencies in the 21st century. The model's fit was poor through the 1970s and early 1990s and has risen almost monotonically since. They make the case that better monetary policy is the reason, because credible inflation targeting removed the scope for self-fulfilling expectations.

But fit is not forecasting, and the difference is the whole puzzle. Engel and Wu measure fit — t- and F-statistics, R-squareds — which asks how well a model describes exchange rate movements when you already have the data. Meese and Rogoff asked something harder: whether a model beats a random walk at predicting movements it has not seen. A model can do the first far better than it used to and still not do the second.

The out-of-sample evidence is real but narrower. Tanya Molodtsova and David Papell (2009) find short-horizon predictability against the dollar for 11 of 12 currencies using Taylor rule fundamentals, and report that the evidence is much stronger with those than with conventional interest rate, purchasing power parity or monetary models. Short-horizon, Taylor rule, 11 of 12 — that is the claim the literature supports, and it is a good deal more specific than "the models work now."

The caution comes from the same authors. In earlier work on the dollar at 12-, 36- and 60-month horizons, Engel and Wu (2021) found apparently strong forecasting power from the exchange rate's own level — and then, correcting for small-sample bias by simulation, found little evidence to reject a random walk. The cause was near-spurious correlation between a persistent predictor and a long horizon, and they warn the same problem may arise for other persistent variables. Long-horizon FX predictability is exactly where a result is most likely to be an artefact.

What is left standing is worth having: the models did not get smarter, the relationships they were built to capture got more stable, and that shows up clearly in fit and modestly in short-horizon forecasting.

What This Means for the Framework

The 4xForecaster framework reads structural forces at a medium-term horizon — weeks to months — and daily updates do not mean daily trading signals. They mean a continuously updated assessment of where those forces are pointing. That horizon is a design choice, not something the papers above establish. Molodtsova and Papell's out-of-sample result is short-horizon, and Engel and Wu's own caution is that the long end is where apparent predictability is least trustworthy.

When the rate differential layer shows a widening spread favouring the dollar, what that describes is a structural force the literature finds better behaved than it was in the 1980s — not a forecast, and not a horizon anyone here has measured. No lag is quoted in this piece because none of the papers above establishes one for this framework's signals.

The puzzle is not solved so much as better understood: the world the models were built for arrived after they were written off. Which signals to weight, and how to read them in sequence, is what the framework article covers in full, and what the rate differentials article unpacks for the most foundational signal layer.