Do AI-Based Forecasting Systems Improve Predictability in Foreign Exchange Markets Compared to Classical Econometric Models?

This paper presents a comparative review of the literature on foreign exchange rate forecasting, with focus on whether artificial intelligence methods can improve prediction results when compared with classical econometric models. Recent research has moved from theory-based statistical models toward data-based methods that are able to identify nonlinear patterns and use both numerical market data and information from financial texts. However, the results of these studies are difficult to compare directly because they use different currency pairs, time periods, forecasting horizons, input variables, and evaluation methods. This makes it harder to draw a clear conclusion about how large and stable the reported improvements are.

The review examines fifteen peer-reviewed studies indexed in Web of Science, most published between 2019 and 2025. The studies are grouped into four categories: classical econometric models, machine learning models, deep learning models, and approaches connected with large language models. Based on document review and qualitative comparison, the paper analyzes whether artificial intelligence methods perform better than econometric benchmarks in out-of-sample forecasting. It also looks at the main design choices that may explain better forecasting results.

The evidence reviewed in this paper suggests that artificial intelligence models often achieve better results than classical econometric models on common forecasting measures such as root mean squared error, mean absolute error, and directional accuracy. Deep learning models, especially long short-term memory variants and attention-based models, tend to show stronger performance in several studies. The review also shows that models using financial text may add useful information that is difficult to include in standard econometric models. The paper brings together findings from different strands of the literature and shows where the improvements are more visible. At the same time, it points out several remaining problems, especially model interpretation, stability across different market conditions, and practical use in real forecasting settings.