AI and Exchange Rate Predictability
This paper demonstrates that generative AI models can effectively resolve the Meese-Rogoff exchange rate disconnect puzzle by forecasting currency returns based on economic fundamentals, achieving a Sharpe ratio exceeding 0.7 through a strategy that leverages the Taylor rule framework as a key predictive mechanism.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Imagine the global economy as a massive, chaotic ocean where currencies are like boats. For decades, economists have been trying to predict which boat will rise and which will sink by looking at the weather reports—the "economic fundamentals" like inflation, jobs, and growth. But there's a famous mystery called the "exchange rate disconnect." It's like trying to forecast a storm by reading a weather map, only to find that the boats move completely randomly, ignoring the map entirely. A simple guess (a "random walk") often works better than complex math. This is frustrating because if we can't predict how money moves, it's hard to understand the world's biggest financial market. Enter Artificial Intelligence (AI), specifically a new kind called "Large Language Models." Think of these not as calculators, but as super-smart, tireless interns who can read thousands of news headlines and economic reports in a split second, connecting dots that human brains might miss. The big question is: Can this digital brain finally crack the code of the ocean, or is it just guessing?
This paper, written by Amin Izadyar, asks exactly that. The author uses AI models like ChatGPT and DeepSeek to act as a financial analyst for ten major currencies (like the US Dollar, Euro, and Japanese Yen). Instead of feeding the AI complex math equations, the author gives it a simple job: read the latest economic news (like a GDP report or an inflation number), compare it to what experts predicted, and decide if the news is good (the currency will get stronger), bad (it will get weaker), or neutral. The AI then spits out a "directional signal." By collecting thousands of these signals over time, the author builds a scorecard called the AIFX index, which measures the net "strength" of a currency based on its economic news.
The results are surprisingly promising. The paper finds that this AI-generated scorecard can actually predict currency returns. A simple trading strategy that buys currencies with strong AI scores and sells those with weak scores generates a profit (measured by a "Sharpe ratio") of over 0.7 per year. To put that in perspective, this is a solid return that beats just guessing, and it remains profitable even after accounting for other known ways traders make money. The AI seems to have found a hidden link between economic news and currency prices that traditional models missed.
However, the author is very careful not to claim the AI is magic or that it has "solved" the puzzle forever. A major worry with AI is "look-ahead bias"—the fear that the AI is just using information from its training data rather than actually reasoning. To prove this isn't happening, the author runs four clever tests. First, they ask the AI to guess the year a news story was published without being told the date; the AI fails miserably, guessing the wrong year most of the time. Second, they compare an older AI model (trained on data up to 2021) with a newer one (trained up to 2023); both perform similarly, suggesting the newer one isn't just "remembering" recent events. Third, they check if the AI has memorized the historical relationship between things like inflation and currency prices; it hasn't. Finally, they create a "pure hindsight" portfolio based on what the AI thinks it remembers about past price movements, and find that the AI's actual trading strategy works in the opposite direction of mere memory. These tests suggest the AI is genuinely reasoning about the news, not just reciting a script.
Digging deeper, the paper discovers why the AI works. It turns out the AI is most successful when looking at three specific types of news: Inflation data, Employment data, and Broad economic activity indicators. These are the exact same variables that central banks use in a famous formula called the Taylor Rule to decide interest rates. The AI seems to intuitively understand that if a country has high inflation or low unemployment, the central bank will likely raise interest rates, making that currency more attractive. Interestingly, the AI's success is driven mostly by positive news (good economic data). While bad news causes an immediate, sharp drop in currency value (which the market digests instantly), good news leads to a slower, more gradual rise, giving the AI more time to predict future gains.
In short, the paper suggests that Artificial Intelligence can act as a highly effective financial analyst, uncovering a real connection between economic fundamentals and exchange rates that has been hiding in plain sight. It doesn't mean the market is perfectly predictable, but it does suggest that with the right tools, we can finally start reading the weather map correctly. The findings are robust and backed by rigorous testing, but the author notes that this is a starting point for understanding a complex system, not a final answer to every mystery in finance.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.