A 1983 study tested whether bank FX forecasts could beat a coin flip.
They lost.
Forty years later, the academic literature has not changed its mind: currency forecasters are, on average, worse than predicting "tomorrow looks like today." So why does every major bank still publish them?
The Meese-Rogoff result that broke the field
In 1983, economists Richard Meese and Kenneth Rogoff published a paper in the Journal of International Economics with a quiet thermonuclear conclusion. They tested all the major exchange rate models of the era — purchasing power parity, the monetary model, the portfolio-balance model — against a "random walk" forecast that said "tomorrow's exchange rate equals today's exchange rate."
The random walk won. At every horizon from 1 month to 12 months, on every major currency pair, the simple "no change" forecast had a lower mean squared error than every structural model the field had built.
The result has been replicated dozens of times over four decades, with new data, new models, machine learning, and orders of magnitude more computing power. The random walk keeps winning at short horizons.
The bank track record nobody publishes
Bloomberg surveys major bank FX desks every quarter for their year-end USD/EUR, USD/JPY, and GBP/USD forecasts. The aggregated consensus is then compared to the actual close.
The 2023 consensus for end-2024 EUR/USD: 1.12. Actual close: 1.035. Miss: 8%.
The 2022 consensus for end-2023 USD/JPY: 130. Actual close: 141. Miss: 8%.
The 2021 consensus for end-2022 GBP/USD: 1.42. Actual close: 1.21. Miss: 17%.
These are the consensuses of dozens of professional FX strategists with multi-million-dollar research budgets. The average miss across major pairs over the last decade is roughly 7%. The random-walk forecast — assume the current spot is the year-end spot — averages a smaller error.
If you'd done nothing, you'd have done better than the experts.
Why FX is uniquely hard
Three reasons FX forecasting fails worse than most other markets.
One: currencies are relative prices. Stocks have a fundamental value linked to discounted cash flows. Currencies don't. The "price" of a dollar in euros depends entirely on what people are willing to swap for it right now, which is the very thing being forecast.
Two: central bank action dominates fundamentals. A single Fed press conference can move EUR/USD more than three months of trade-balance data. Models built on fundamentals get repeatedly overwritten by surprise policy moves nobody can predict.
Three: noise traders matter. Currency markets are the deepest in the world — $7.5 trillion turns over every day — but that depth is dominated by hedging flows, corporate transactions, and short-term speculation. The "informed" signal in the daily price is genuinely small.
What the carry trade tells you
If forecasts were any good, the carry trade — borrowing in low-yielding currencies and lending in high-yielding ones — should be unprofitable. The market should adjust spot rates so that the high-yield currency depreciates enough to wipe out the interest differential.
It hasn't. The carry trade has generated positive returns over decades, with periodic blowups. This is known as the "forward premium puzzle," and it has been unsolved for the entire history of the modern FX market.
Either the entire global FX market is mispriced in a persistent and predictable way for forty years — or the models we use to value currencies are wrong. Almost every economist now believes the second.
What actually works
A small handful of approaches have positive track records.
Real interest rate differentials, over horizons of 3+ years. The currency with the higher real rate tends to outperform — but the noise drowns out the signal over any window short enough to act on.
Trend-following. Currencies trend. Systematic trend-followers in FX have positive Sharpe ratios over multi-decade windows. But individual moves are dominated by mean-reversion noise.
Crisis prediction. Currency crises (Mexican peso 1994, Asian crisis 1997, ruble 2014, Argentina nearly every other year) have observable precursors: rising current account deficits, fixed-rate regimes under pressure, capital flight indicators. These are predictable. The exact timing is not.
Why prediction markets do better here
Prediction markets do not "forecast" currencies in the traditional sense. They price specific binary events: "Will USD/EUR close above 1.10 on June 30?"
That framing forces traders to confront probabilities, not point estimates. A bank can publish a point forecast of 1.08 and never have to live with it. A market trader who buys "above 1.10" at 35 cents and sees it close at 1.07 loses real money.
Across the major USD pairs, the implied probabilities from FX-options markets — the closest existing analog to a prediction market — outperform bank consensus forecasts in calibration tests. Markets are sharper than committees.
How Juno fits in
Juno is building a prediction market for global retail traders. FX is one of the most natural use cases: clear, recurring, internationally relevant questions, with end dates that resolve unambiguously.
"Will the Fed cut rates by 50bps before September?" "Will USD/EUR close below 1.05 by year-end?" "Will the next BoJ meeting raise rates?" Each becomes a contract. The price tells you what the crowd actually believes — not what the bank's PR-approved year-end note says.
The 1983 lesson
Meese and Rogoff didn't set out to embarrass an industry. They set out to compare models. The result was that the industry's best work couldn't beat doing nothing.
Forty years on, the FX forecasting industry still exists. The forecasts still get published. They still get quoted. They still don't beat the random walk.
The next time a bank tells you where USD/EUR will be in twelve months, ask them a different question: where would it be if they took the year off?