The prediction of exchange rates has been a longstanding challenge in financial economics. Exchange rates play a pivotal role in international trade, investment decisions, and economic policy. The volatility and unpredictability of currency movements can have profound implications for businesses, governments, and financial institutions. Therefore, accurate forecasting of exchange rates is essential for making informed decisions in the global financial markets. Traditional models of exchange rate forecasting, which rely heavily on economic theories such as purchasing power parity (PPP) and interest rate parity (IRP), have faced significant limitations in capturing the complexities of currency movements in real-time markets. While these models provide valuable insights into long-term trends, they often fail to predict short-term fluctuations due to their reliance on simplified assumptions and linear relationships between macroeconomic variables. As a result, forecasting exchange rates has remained an area where there is a high degree of uncertainty and error.

Over the past few decades, the rapid advancement of Artificial Intelligence (AI) and Machine Learning (ML) techniques has introduced new approaches to forecasting complex financial phenomena like exchange rates. AI and ML models are capable of processing vast amounts of data, identifying non-linear patterns, and learning from historical trends without the need for explicit assumptions. Unlike traditional econometric models, which typically rely on predefined equations and theoretical relationships, AI and ML models are data-driven and can adapt to dynamic market conditions. This has led to a growing interest in applying machine learning algorithms, such as neural networks, decision trees, and support vector machines, to predict exchange rate movements. Since 2020, the application of advanced neural architectures such as Long Short-Term Memory (LSTM) networks and Convolutional Neural Networks (CNNs) in exchange rate forecasting has expanded rapidly. These models have proven particularly effective in volatile environments where nonlinear relationships dominate and macroeconomic signals are complex to interpret. For instance, Abedin et al. (1) demonstrated that macroeconomically-augmented LSTM models deliver significantly improved predictive performance for EUR/USD exchange rates, especially during periods marked by sudden economic events such as interest rate hikes or inflationary shocks. Similarly, Dritsas and Trigka (2, 3) showed that deep LSTM ensemble models consistently outperform both traditional econometric approaches and shallow neural networks, particularly in multi-horizon forecasting contexts. These developments highlight a major shift in recent literature toward more flexible, adaptive, and non-linear modeling strategies to capture the complex dynamics of currency markets.

These algorithms have been shown to outperform conventional models, especially in volatile market environments where traditional models struggle to adapt quickly. Despite the promising results of AI and ML models in exchange rate forecasting, the question remains whether these models fully incorporate macroeconomic factors that significantly influence exchange rates. Factors such as interest rates, inflation, and unemployment are known to play a fundamental role in determining the long-term direction of exchange rates. However, the impact of these macroeconomic indicators on the performance of AI and ML models, particularly in short-term forecasting, is not well understood. While previous research has focused on improving the statistical accuracy of these models, there has been limited attention paid to their economic interpretability and their ability to integrate macroeconomic variables in a meaningful way. This gap in the literature represents a crucial area of inquiry, as understanding the relationship between forecast errors and macroeconomic factors can provide deeper insights into the strengths and limitations of AI/ML models for practical financial forecasting.

The primary aim of this study is to conduct a detailed economic analysis of the performance of AI and ML models in exchange rate forecasting by examining the relationship between forecast inaccuracy and key macroeconomic factors, including interest rates, unemployment, and inflation. While previous studies have evaluated the accuracy of machine learning models in forecasting exchange rates, this research takes a novel approach by focusing specifically on the economic context in which these models operate. The study uses a monthly time series dataset spanning ten years, from January 2014 to December 2024, to capture a range of economic cycles, including the recovery from the global financial crisis, the COVID-19 pandemic, and the recent inflationary and interest rate shocks in the United States and Europe. The EUR/USD exchange rate is chosen as the focus of this analysis due to its prominence in global financial markets and its sensitivity to macroeconomic factors from both the U.S. and the Eurozone.

The study employs regression models to explore how macroeconomic indicators influence the forecast errors of AI and ML models. By analyzing the relationship between forecast inaccuracy and variables such as GDP growth, inflation, interest rates, and unemployment, the study aims to fill the gap in the existing literature, which largely overlooks the role of macroeconomic factors in the performance of these models. In addition to examining the direct influence of these indicators on forecast errors, the study also investigates the overall effectiveness of AI/ML models compared to traditional econometric techniques in terms of both forecasting accuracy and economic interpretation. While statistical accuracy, measured by indicators like the root mean square error (RMSE) and mean absolute percentage error (MAPE), remains a key focus, this study emphasizes the economic relevance of forecast errors, particularly in the context of real-world financial decision-making.

The methodology of this research is designed to provide robust results by employing a variety of statistical techniques, including multiple regression analysis, correlation analysis, and ANOVA tests. The use of a comprehensive dataset spanning over a decade allows the study to capture a range of economic conditions, from periods of relative stability to those marked by significant volatility and macroeconomic shocks. This enables a more nuanced understanding of the performance of AI and ML models in different economic contexts. The study also incorporates control variables to account for confounding factors that may influence both macroeconomic indicators and model performance. For example, binary indicators are used to account for significant economic events, such as financial crises or policy changes, that could have an impact on exchange rate movements and forecast errors.

A key aspect of this study is the comparison between AI/ML models and traditional econometric models. While machine learning algorithms have demonstrated superior performance in terms of predictive accuracy, their economic interpretability has often been questioned. Traditional models, on the other hand, may lack the flexibility and adaptive capabilities of AI/ML models but offer clearer theoretical underpinnings. By comparing the two approaches, this study seeks to determine whether AI/ML models can provide not only better forecasts but also insights into the underlying economic dynamics that drive exchange rate movements. This comparison is crucial for understanding how AI/ML models can complement or even replace traditional models in financial forecasting.

In terms of findings, this study anticipates that AI and ML models will outperform traditional econometric models in terms of accuracy, especially in forecasting short-term fluctuations in the EUR/USD exchange rate. However, it is expected that macroeconomic factors, such as the Eurozone interest rate, will show some level of correlation with forecast errors, albeit weak. This suggests that while AI/ML models can capture complex relationships in historical exchange rate data, they may still be sensitive to significant macroeconomic events that cannot be fully captured by historical price movements alone. The study’s results are expected to provide valuable insights into the strengths and limitations of AI/ML models, especially in terms of their ability to integrate macroeconomic variables and provide reliable forecasts in times of economic instability.

The contribution of this study lies not only in its exploration of the effectiveness of AI and ML models for exchange rate forecasting but also in its economic analysis of forecast errors. By examining how these errors relate to macroeconomic indicators, the study provides a more comprehensive understanding of the performance of AI/ML models in real-world financial forecasting. The findings of this research have important implications for financial institutions, policymakers, and investors who rely on exchange rate forecasts to make strategic decisions. Moreover, the study highlights the potential of AI and ML models to revolutionize financial forecasting by offering data-driven insights that go beyond traditional economic theories.

In sum, this research contributes to both the empirical and theoretical literature by evaluating the performance of AI/ML models in exchange rate forecasting through the lens of macroeconomic integration. By bridging data-driven techniques and economic reasoning, the study offers a dual perspective: assessing predictive accuracy and uncovering the role of fundamental indicators in explaining model errors. This dual approach is particularly relevant for enhancing the interpretability and policy relevance of machine learning applications in finance.

Exchange rate forecasting is a critical component of financial markets and international trade, serving as an essential tool for businesses, investors, and policymakers alike. The ability to predict exchange rate movements accurately has far-reaching implications. For businesses, it enables the development of effective currency hedging strategies, protecting against unexpected fluctuations in exchange rates that could otherwise undermine profitability (4). For investors, exchange rate forecasting plays a pivotal role in managing foreign exchange risk, allowing them to make well-informed decisions regarding portfolio allocation and investment strategies (5). Policymakers, too, rely on accurate exchange rate forecasts to make strategic decisions about monetary policy and trade agreements, as the value of a currency impacts national inflation, export and import prices, and overall economic growth (6, 7).

The foreign exchange market, or forex, is the largest and most liquid market globally, with a daily trading volume exceeding $6 trillion (45). Given the market’s size and complexity, exchange rate forecasting is central to financial risk management practices. The movement of currencies directly affects trade balances, investment flows, inflation rates, and even global capital flows. For instance, fluctuations in exchange rates can significantly alter the competitiveness of a country’s goods and services in global markets, thereby influencing trade dynamics (8). As such, exchange rate forecasts provide essential insights for anticipating economic conditions and formulating policy decisions (9).

Historically, the approaches to exchange rate forecasting were dominated by traditional econometric models. Among the most widely used were the Autoregressive Integrated Moving Average (ARIMA), Vector Autoregressive (VAR) models, and the random walk model. These models predominantly relied on statistical techniques and historical data to predict future exchange rate movements. ARIMA, for example, is a time-series model that uses past data points to forecast future values based on their observed relationships (10). Similarly, VAR models captured the interdependencies among multiple time-series variables, making them useful for examining how different economic factors such as interest rates and inflation affect exchange rates (11). The random walk model, which assumes that future exchange rate movements are unpredictable and do not follow any systematic trend, was also frequently employed in forecasting exercises (12).

However, while these traditional econometric models provided some valuable insights, they often had notable limitations. One of the primary drawbacks was their reliance on linear assumptions, which limited their ability to capture the complex, non-linear relationships that are prevalent in financial markets. Plakandaras and Papadimitriou (13) highlight that exchange rates are influenced by a wide variety of factors, including geopolitical events, speculative activity, and market sentiment, which cannot always be captured by linear models. Furthermore, these econometric models struggled to account for the sharp, often unpredictable shifts in the market that are characteristic of the forex market. For instance, during periods of financial crises or political instability, exchange rates can experience sudden and extreme movements that linear models, such as ARIMA or VAR, are ill-equipped to predict (14, 46).

Moreover, the random walk model, while useful in suggesting that exchange rate predictions are difficult in the long run, failed to provide much insight into short-term movements or volatile market conditions (7). This limitation of traditional models such as the random walk was notably illustrated in the influential work of Meese and Rogoff (12), who concluded that no macroeconomic model outperformed a naïve random walk in out-of-sample forecasting. However, more recent studies using machine learning have challenged this view. For instance, Küçük (15) showed that LSTM and CNN architectures significantly outperformed ARIMA and VAR models in predicting exchange rate movements by effectively capturing nonlinear dependencies in macroeconomic variables like inflation and interest rates. Similarly, Ciganovic (16) demonstrated that advanced AI models incorporating labor market dynamics, such as unemployment duration, provided enhanced forecasting accuracy across different economic regimes. Furthermore, Mirza (17) emphasized the role of explainable AI techniques to bridge the interpretability gap between machine learning models and traditional econometric approaches, thereby improving trust and usability in financial forecasting. These findings signal a growing consensus that AI-driven models not only surpass traditional methods in predictive power but also offer better adaptability to structural changes and macroeconomic shocks. As the forex market is influenced by a multitude of factors, many of which interact in highly complex and dynamic ways, the random walk model proved insufficient for forecasting exchange rates during periods of high volatility or when market shocks occur.

To address these limitations, newer approaches, particularly those based on machine learning (ML) and artificial intelligence (AI), have gained popularity. These newer models can account for the complexity and non-linearity inherent in financial data by leveraging vast amounts of historical and real-time data. For instance, models such as support vector machines (SVM), neural networks, and long short-term memory (LSTM) networks have shown significant promise in capturing patterns and trends that traditional econometric models often missed (2, 3, 47). These machine learning models are capable of incorporating not only economic variables like inflation rates and GDP growth but also unstructured data such as news articles, social media sentiment, and even geopolitical events, providing a more holistic view of the factors driving exchange rate movements (18, 19).

The SVM model, for example, has been effective in classifying exchange rate trends and predicting directional movements based on historical data (20). Similarly, neural networks have demonstrated their ability to model complex non-linear relationships and learn from vast datasets, offering a more adaptable and accurate approach compared to traditional models (21, 22). Moreover, the advent of LSTM networks has significantly enhanced exchange rate forecasting by allowing models to capture temporal dependencies and long-term trends in time-series data, making them well-suited for financial forecasting (23, 24).

Despite the promise of AI and ML-based models, they are not without their challenges. One of the primary concerns is the need for large amounts of high-quality data to train these models effectively. Incomplete, noisy, or biased data can lead to inaccurate predictions, and overfitting—where a model fits the training data too closely—remains a persistent issue in the use of machine learning for financial forecasting (25, 48). Furthermore, the lack of interpretability in many AI and ML models, particularly deep learning techniques, is another challenge. The so-called “black-box” nature of models like neural networks and LSTM networks makes it difficult for analysts to understand how a model is arriving at its predictions, which is a significant concern in financial forecasting where transparency and accountability are crucial (26).

Moreover, while these newer techniques have been successful in capturing the complexities of exchange rate movements, they still face limitations when it comes to incorporating macroeconomic fundamentals such as interest rates, inflation, and unemployment. Traditional econometric models were explicitly designed to model these macroeconomic variables and their direct impact on exchange rates, but the ability of ML and AI models to explicitly incorporate these variables and their relationships with exchange rates is still an area of ongoing research (1). Research shows that combining AI and ML techniques with macroeconomic fundamentals could enhance forecasting accuracy, as AI models can detect the complex, non-linear relationships that traditional econometric models often overlook (27).

The challenge of accurately forecasting exchange rates is further compounded by the inherent uncertainty and volatility of the forex market. Political instability, economic crises, natural disasters, and global pandemics all introduce unpredictable fluctuations in exchange rates, rendering even the most sophisticated forecasting models less reliable in the short term (28). As such, forecasters must determine how and when to incorporate external factors, such as geopolitical events or policy changes, into their models. Some recent advancements, such as the use of reinforcement learning and deep learning algorithms, have been explored as solutions to these challenges, allowing models to adapt to unexpected market shocks and improve their robustness in volatile environments (49). However, these models also face challenges in processing large, complex datasets and require careful tuning to avoid overfitting.

The rise of machine learning (ML) and artificial intelligence (AI) has brought forward new forecasting techniques that can handle the complexities and non-linearity of exchange rate data. These newer AI/ML techniques, such as Support Vector Machines (SVM), Random Forests, Neural Networks, and Long Short-Term Memory (LSTM) networks, have been found to outperform traditional econometric models by effectively capturing patterns and trends that the older models missed (2, 3, 47). AI and ML models can incorporate vast amounts of historical and real-time data, such as macroeconomic indicators, sentiment analysis, and even financial news, which makes them adept at forecasting in dynamic and volatile market conditions.

AI/ML models allow for greater flexibility by dynamically adapting to new data without the stringent assumptions required by traditional models, such as stationarity. This is particularly advantageous during periods of high market volatility, as AI models can continuously adjust to changing market conditions without being restricted by prior assumptions about data behavior (29). However, despite their promise, these models are not without challenges, including concerns around overfitting, the need for large amounts of high-quality data, and interpretability issues (17).

Moreover, the increased reliance on macroeconomic variables—such as interest rates, inflation rates, GDP growth, and unemployment—remains pivotal to exchange rate dynamics. While traditional econometric models emphasized these variables as key predictors of exchange rates, AI/ML models typically rely on historical relationships between variables and may not always explicitly model the intricate interactions between exchange rates and macroeconomic fundamentals. Despite this, research suggests that combining AI/ML techniques with macroeconomic fundamentals can improve forecast accuracy, as AI models are better equipped to detect non-linear relationships that traditional models may overlook (1).

One of the main challenges in exchange rate forecasting is the inherent uncertainty and volatility of currency markets. Factors such as political instability, natural disasters, and global economic events like financial crises or pandemics introduce unpredictable fluctuations that make forecasting difficult. These external factors complicate the forecasting process, as it is often unclear how or when they should be incorporated into forecasting models (28). However, recent advancements in AI—specifically reinforcement learning and deep learning algorithms—have shown potential in adapting to such shocks and improving model robustness in volatile environments (49).

Despite the promising capabilities of AI and ML models, some challenges remain. While traditional models generally perform well during periods of economic stability, they often fall short when forecasting in the face of unexpected economic shocks or market disruptions. The adaptability of AI/ML models to these unexpected events could be further enhanced by incorporating real-time data sources, such as financial news, social media sentiment, and geopolitical events (22). Furthermore, it has been shown that combining AI/ML models with hybrid approaches that integrate macroeconomic variables and econometric techniques can enhance forecasting accuracy during times of high volatility (30).

Macroeconomic variables play a central role in exchange rate forecasting. Interest rates, inflation, GDP growth, and unemployment rates have long been established as significant drivers of exchange rate movements. In traditional econometric models, these variables were considered the primary predictors of exchange rate changes. For example, the Interest Rate Parity (IRP) theory, which states that exchange rates should adjust based on interest rate differentials between two countries, has been foundational in understanding how interest rates affect exchange rates (31). Likewise, the Purchasing Power Parity (PPP) theory posits that exchange rates should adjust to equalize the price levels between two countries (32, 33). However, these traditional models often fail to account for the complexities and non-linear dynamics of real-world financial markets. For example, exchange rates are often influenced by factors such as market sentiment, political instability, and speculative behavior—factors that are difficult to quantify or model directly in econometric frameworks (34, 35). As a result, more dynamic forecasting approaches, such as AI/ML models, have been developed to capture these complex, non-linear relationships. While PPP and IRP provide foundational frameworks for explaining exchange rate movements, more advanced models such as the Behavioral Equilibrium Exchange Rate (BEER) and the Fundamental Equilibrium Exchange Rate (FEER) offer improved empirical performance, especially for medium-term alignment. Clark and MacDonald (50) developed the BEER model to account for productivity differentials, interest rate spreads, and net foreign assets, providing a behavioral benchmark for exchange rate misalignments. Similarly, Williamson (51) formulated the FEER model to identify exchange rates consistent with internal and external balance, offering useful policy guidance. These frameworks complement AI/ML approaches by embedding structural economic reasoning into empirical analysis.

AI/ML models are capable of incorporating a wider range of macroeconomic data, financial data, and unstructured data sources such as news sentiment and market opinions. These models are also better equipped to detect patterns in historical data and assess how they relate to future exchange rate movements. Studies have shown that incorporating macroeconomic indicators like interest rate differentials, inflation, and GDP growth into AI/ML models can significantly improve forecast accuracy (52). These improvements in forecasting accuracy are particularly important in volatile market environments, where traditional models often fail to capture the sudden changes in market sentiment or economic conditions.

Hybrid models that combine the strengths of traditional econometric techniques with the adaptability and flexibility of AI/ML models are gaining traction in exchange rate forecasting. These hybrid models seek to combine the theoretical underpinnings of econometric models, such as the macroeconomic variables that drive exchange rates, with the empirical, data-driven patterns captured by AI and ML. For example, integrating ARIMA or Vector Autoregression (VAR) models with deep learning architectures such as LSTM or CNNs can allow researchers to capture both the macroeconomic relationships between variables and the non-linear, complex patterns in the data (36). This approach has proven effective in capturing both short-term and long-term fluctuations in exchange rates, allowing for more accurate predictions. Furthermore, multi-source data integration is an important aspect of hybrid models. Traditional econometric models typically relied on structured data, such as historical exchange rates and macroeconomic indicators. However, the integration of unstructured data—such as financial news, social media sentiment, and geopolitical events—into AI/ML models has become increasingly important in improving the robustness and accuracy of exchange rate forecasts (37). By incorporating a broader range of data sources, hybrid models can more effectively capture the dynamic nature of financial markets and improve their predictive performance.

While AI and ML models offer numerous advantages over traditional econometric approaches, they are not without their challenges. One of the most significant challenges in AI/ML modeling is the risk of overfitting. These models have a large number of parameters, which increases the likelihood of fitting noise or irrelevant patterns in the data. Overfitting occurs when a model performs well on the training dataset but fails to generalize effectively to new, unseen data. This is a particular concern in the financial markets, where data is often noisy and market conditions change rapidly (25). Regularization techniques such as dropout and L2 regularization can help mitigate overfitting, but finding the right balance between model complexity and generalization remains a critical challenge (22).

Another challenge in AI/ML modeling is interpretability. Unlike traditional econometric models, which are relatively straightforward and based on economic theory, AI and ML models are often referred to as “black boxes” due to their complexity and lack of transparency. This lack of interpretability can be a significant barrier to the adoption of AI/ML models in financial forecasting, as decision-makers often need to understand the rationale behind the predictions made by the model (26). Researchers have proposed methods to improve the interpretability of AI/ML models, such as LIME (Local Interpretable Model-agnostic Explanations), but these methods are still in their early stages and have not yet been widely adopted in financial forecasting (38).

The primary objective of this study is to perform a comprehensive economic analysis of the effectiveness of AI and ML models in exchange rate forecasting by examining the relationship between the forecast inaccuracy of these models and important macroeconomic factors, including GDP growth, interest rates, and inflation. This section presents the methodological framework, detailing the research design, variables, data collection, and analysis techniques.

The primary objective of this study was to assess the performance of AI and ML models in forecasting the EUR/USD exchange rate, with a particular focus on evaluating the relationship between forecast inaccuracy and key macroeconomic factors such as interest rates, inflation, GDP growth, and unemployment rates. The forecasting model was based on lagged linear regression, using Microsoft Excel’s Data Analysis ToolPak to generate predictions for the EUR/USD exchange rate.

The findings of this study reinforce the growing consensus on the effectiveness of AI and ML models in exchange rate forecasting. Among the models tested, the LSTM network exhibited the highest predictive accuracy (R² = 0.9234), outperforming both the MLP and Random Forest models. This demonstrates the ability of deep learning architectures to capture sequential and nonlinear dependencies in exchange rate data.

However, the analysis also reveals that the forecast errors of these models are only weakly correlated with traditional macroeconomic indicators. For instance, only the Eurozone interest rate demonstrated a statistically significant, albeit weak, relationship with forecast errors. Other variables—such as inflation, GDP growth, and unemployment—showed no robust explanatory power. This finding suggests that short-term exchange rate fluctuations may be driven more by market dynamics, speculative behavior, or high-frequency technical signals than by macroeconomic fundamentals.

Traditional linear modeling techniques may fail to capture the complexity of macroeconomic dynamics. According to Sarno and Taylor (33), exchange rate movements often display nonlinear patterns and regime-dependent behavior, especially during periods of macroeconomic instability. Similarly, Krolzig (54) emphasizes the value of Markov Switching models in capturing structural shifts in economic relationships. Moreover, studies such as Kim and Nelson (55) have shown that macroeconomic effects on exchange rates are often asymmetric and vary across different market conditions. These findings suggest that purely linear regression frameworks might underestimate the impact of macro variables, particularly when their relationships with exchange rates are conditional or time-varying.

This aligns with recent literature, such as Plakandaras et al. (42), who argues that macroeconomic indicators often lose their short-term predictive value during periods of structural breaks or policy regime shifts. Likewise, Mazumder and Chatterjee (43) emphasizes that raw macro indicators may require complex transformation, interaction modeling, or embedding within neural structures to gain explanatory power. The weak correlation observed in this study may thus stem from modeling constraints rather than the irrelevance of fundamentals.

To improve forecast sensitivity, future modeling efforts should consider richer input pipelines—such as engineered macroeconomic features, temporal segmentation (e.g., pre- and post-crisis), and integration of real-time alternative data (e.g., sentiment indices, financial news). Additionally, explainable AI (XAI) tools could help identify the latent influence of economic variables in deep learning models, enhancing both performance and interpretability.

In conclusion, this study demonstrates that while AI and ML models like LSTM offer high accuracy in forecasting currency movements, their integration with macroeconomic knowledge remains a methodological challenge. Enhancing these models with feature-engineered, time-aware macroeconomic data and interpretable structures could bridge the gap between statistical precision and economic insight.

This study presents an in-depth evaluation of the performance of Artificial Intelligence (AI) and Machine Learning (ML) models in forecasting exchange rates, with a particular emphasis on the EUR/USD currency pair. By implementing three complementary architectures—Multilayer Perceptron (MLP), Random Forest (RF), and Long Short-Term Memory (LSTM)—the analysis assessed both predictive accuracy and the macroeconomic sensitivity of forecast errors. The LSTM model achieved the highest performance (R² = 0.9234), followed by RF and MLP, confirming the capacity of deep learning to capture complex short-term exchange rate dynamics based on historical data.

However, despite these promising results, the study uncovered only limited explanatory power of macroeconomic indicators—such as inflation, interest rates, and unemployment—over forecast errors. This suggests that while AI/ML models effectively detect patterns in past prices, they may struggle to integrate economic fundamentals unless those inputs are properly transformed, lagged, or contextualized. While this study employed a multiple linear regression framework to explore the relationship between forecast errors and macroeconomic indicators, it is worth noting that alternative modeling approaches—such as nonlinear regressions, structural models, or regime-switching techniques (e.g., Markov Switching Models)—may offer improved statistical significance. […] Nonetheless, these avenues represent promising directions for future research aimed at deepening the integration between economic theory and machine learning.

To address this, the study incorporated model interpretability tools, including forecast comparison plots, variable importance charts, and loss curves. These diagnostics enhanced transparency and supported the assessment of each model’s behavior.

Recent studies have emphasized the importance of explainability in financial machine learning. Ribeiro et al. (38) introduced LIME as a way to interpret model predictions locally, while Lundberg and Lee (56) proposed SHAP to compute consistent and theoretically grounded feature attributions. In the context of macro-financial forecasting, Abedin et al. (1) demonstrated that integrating SHAP explanations into deep learning pipelines can uncover latent economic signals and improve trust in model decisions. The use of such tools is particularly relevant in complex models like LSTM and Random Forests, where the internal decision-making process is often opaque. Therefore, XAI methods play a critical role in bridging the gap between statistical accuracy and economic interpretability.

Additionally, the results point toward the future potential of hybrid and explainable AI (XAI) approaches. As suggested in recent studies (1, 17, 44), combining macroeconomic knowledge with tools such as SHAP or LIME can lead to more transparent, accurate, and economically grounded forecasting systems.

In conclusion, while AI and ML models like MLP, RF, and LSTM show considerable promise in forecasting currency exchange rates, their full potential depends on the development of hybrid frameworks that balance statistical performance with economic interpretability. Bridging this gap will enable the creation of forecasting tools that are not only technically robust but also meaningful and actionable in real-world financial applications such as monetary policy planning, portfolio management, and risk mitigation.