PTCan tomorrow's dollar be predicted?
For 40 years, economists have tried to beat a simple guess: tomorrow's exchange rate will be the same as today's. This research put six artificial intelligence techniques into the contest, with five currencies, day by day, from 2010 to 2023. None erred less than the simple guess. Explained in 10 screens.
Podium by lowest error across the five currencies: the "tomorrow = today" guess came 1st in all of them; LSTM, 2nd in all. *3rd: ARIMA for the yuan, the euro and the pound; N-BEATS for the real; Prophet for the yen.
The hardest guess to beat: "tomorrow = today"
If the market already builds everything known into the price, the next rise or fall is unpredictable. In that case, the best guess for tomorrow is today's price. Economists call this a random walk.
Try to beat it. In each round, say whether the dollar will rise, stay or fall. We count two things: how far off you were on the value and whether you got the direction right. The research measures both.
Illustrative exercise. The series is drawn at random by the computer, with rises and falls left to chance, as the efficient market theory assumes (Fama, 1970). It is not the real exchange rate.
Four decades trying to beat chance
In 1983, Meese and Rogoff showed that the best economic models of the time erred more than "tomorrow = today". Since then, every win over it has come only in specific cases. Artificial intelligence is the latest attempt. Tap the milestones.
Milestones evenly spaced, not to time scale. Studies cited in the dissertation's literature review (chapter 2).
Five currencies, 14 years of daily rates
The Brazilian real, the Chinese yuan, the euro, the pound and the yen, each against the US dollar, from January 2010 to December 2023, from the Bloomberg platform. About 3,300 trading days per currency. The models' task: predict the next day's value.
Bar height: how far the period's highest value was above the lowest, in %. Our calculation from Table 1. The real swung the most. Tap a bar or a currency.
Eight contestants: two yardsticks, one veteran and five newcomers
"Tomorrow = today" and ARIMA, a classic statistical model, are the yardsticks. LSTM, a 1997 neural network that has won in other studies, is the veteran. The five newcomers are recent deep learning techniques (neural networks with many layers). Tap the cards to see how each one works.
No fine-tuning. All the AI models ran with the libraries' default settings (Python, TensorFlow and Darts). The research counts six deep learning techniques: LSTM and the five newcomers.
Nobody beat "tomorrow = today"
For every currency and all three error measures, the simple guess erred least. LSTM came close for four currencies (for the yen, its error was about 10 times larger). The newcomers' errors were from about 7 to over 500 times larger.
Logarithmic scale: each step in height multiplies the error by 10. The number on the bar is the error divided by that of "tomorrow = today", calculated by us from the rounded values in Tables 5 to 9 (approximate when errors are very small).
See the values from the dissertation
The veteran came 2nd; the newcomers, mid-table
The top and bottom never changed: "tomorrow = today" 1st, LSTM 2nd and TimeGAN, from a family of techniques built to generate images, last. In the middle, the newcomers and ARIMA shift places from currency to currency.
Block height: position by RMSE (1st = tallest). The number on top is the position. Tap a block to see the error.
Calling up or down: close to a coin toss
For traders, getting the direction right can matter more than getting the value right. That is why some authors propose this yardstick. The dissertation shows that all the AI models get the direction right far more often than "tomorrow = today".
But there is a catch: "tomorrow = today" never says "up" or "down", so it almost never gets the direction right (0.6% to 5%). The fair comparison is with a coin toss: 50%. The models scored between 40.4% and 54.9%.
Our reading. Direction accuracy (MDA) from Tables 5 to 9. In the dissertation's charts, the test period covers about 700 days. Over that many days, someone guessing with a coin scores between 46% and 54% in 95% of cases. The dissertation does not test whether the models' scores differ from chance.
What changes for those who depend on exchange rates
Exchange rates affect the price of bread, the cost of an import and central bank policy. A model that predicted them well would be worth a lot. The research shows that, in the short run, artificial intelligence is not there yet.
For companies that buy or sell in dollars
- For the next day, today's price is still the best guess.
- Protecting against swings (with hedging contracts) makes more sense than betting on a daily forecast.
For governments and central banks
- Forecasting exchange rates would help anticipate the effects of economic decisions.
- The results call for caution with any tool that promises to get tomorrow's rate right.
For investors
- For the author, getting the direction right is what matters for profit, and the AI models do this better than "tomorrow = today".
- Scores near 50%, however, do not guarantee an edge over chance (screen 08).
For researchers
- Among the newcomers, N-BEATS was the most promising (best average position).
- Next steps: tune the settings, use more data and combine techniques into hybrid models.
Four decades on, the Meese–Rogoff puzzle still stands, now also against artificial intelligence. For the author, given how fast these techniques are advancing, it is worth continuing to try.
What this research does not say
- No fine-tuning. The models ran with default settings. Tuning them, using more data and combining techniques may improve performance. · according to the author
- A single horizon: the next day. Other studies find different results over longer horizons (months to years). · according to the dissertation's review
- Test period not reported. The dissertation does not say which days were used to train the models and which to test them. · our observation
- No statistical test of the differences. The positions come from comparing errors directly. There is no test showing that the differences between models, or between direction scores and 50%, are not due to chance. · our observation
Credits
FRANCISCO, Pedro Paulo Galindo. Alternativas para a previsão da taxa de câmbio: a eficácia dos modelos de deep learning [Alternatives for exchange rate forecasting: the effectiveness of deep learning models]. Master's dissertation (Business Administration) — Mackenzie Presbyterian University, São Paulo, 2024. Advisor: Prof. Dr. Eli Hadad Junior.
Research supported by CAPES (the Brazilian federal agency for graduate education). Daily rates from Bloomberg; models in Python, with the TensorFlow and Darts libraries. This version simplifies the work for a general audience; the technical description of the models and the full tables are in the original text.