These metrics collectively assess the framework’s prediction accuracy, focusing on average errors (MAE) and error variability (MSE and RMSE). Lower values for these metrics indicate better model performance, demonstrating the effectiveness of the hybrid framework in forecasting exchange rates. Table 4 provides notations used in the evaluation methodology.
The experimental results section evaluates the proposed LSTM-DQN model using a comprehensive dataset comprising USD/INR exchange rates and influential economic indicators such as gold, crude oil, and inflation rates. The dataset was pre-processed and normalized for model input. For benchmarking, state-of-the-art models, including Multimodal Fusion-Based LSTM (MF-LSTM), AB-LSTM-GRU, TSMixer, Hybrid LSTM Models, and Two-Layer Stacked LSTM (TLS-LSTM), were used for comparison. MF-LSTM integrates multiple input sources for exchange rate prediction but lacks reinforcement learning. AB-LSTM-GRU combines LSTM and GRU in an AdaBoost framework, ensuring robustness but limited adaptability. TSMixer employs a transformer-based architecture for time-series prediction but struggles with dynamic adjustments. Hybrid LSTM Models leverage macroeconomic and technical indicators for improved accuracy but do not include iterative optimization. TLS-LSTM enhances long-term dependency modeling yet lacks reinforcement-driven adaptability. These methods provide strong baselines, but the proposed model surpasses them in accuracy and robustness. Experiments were conducted in a Python-based environment with TensorFlow and PyTorch libraries on a system equipped with an NVIDIA GPU for accelerated training.
The USD/INR exchange rate data is selected for this study. Its economic importance and inherent volatility make it a clear candidate for analysis of the efficacy of the proposed forecasting framework. Being one of the most traded emerging market currencies, the Indian Rupee (INR) is subject to several macroeconomic indicators, including inflation rates, foreign direct investment (FDI), trade balances, and monetary policies governed by the Reserve Bank of India (RBI) and the Federal Reserve (Fed). Moreover, the behavior of USD/INR is influenced by various global economic factors, movements in oil prices (given India’s status as a large oil importer), and geopolitical events, adding a layer of complexity but also relevance for evaluating the adaptability of deep learning and reinforcement learning based models. This choice also reflects real-world applications of finance where accurate forecasting between the USD and INR currency pairs is essential for investors, multinational corporations, policymakers, and associated forex traders who require hedging from risks to make accurate financial decisions. Using this exchange rate dataset, the study demonstrates the model’s strengths in accommodating complex, non-linear dependencies and enhancing forecasting accuracy in dynamic currency markets. The results demonstrate the superior performance of the proposed model.
In particular, gold prices are closely correlated with INR/USD (Fig. 4). Like INR/USD, it goes up as gold prices go up, but does not continue indefinitely. Since the follow-on distribution does not show a strong linear correlation, there is no unique one-to-one relationship.
The correlation between crude oil prices and the INR/USD exchange rate is illustrated in Fig. 5. It exhibits a positive correlation, indicating that an increase in crude oil prices is accompanied by a corresponding increase in the INR/USD exchange rate. However, this line is not perfectly linear either, and the data points cluster in certain areas and show certain outliers. As such, it means that more than crude oil prices affect the exchange rate, particularly global economic conditions and geopolitical events. Time series, correlation, and regression analyses are methods we can use to analyze this complexity.
Figure 6 illustrates the correlation between crude oil prices and the change in the INR/USD exchange rate. It exhibits a positive correlation, implying that as crude oil prices increase, the INR/USD exchange rate also tends to rise. Indeed, a clear relationship exists, although it could be more linear; the data points cluster in certain areas, and there are outliers. This implies that something other than crude oil prices, such as global economic conditions or geopolitical happenings, drives the exchange rate. This relationship could be further explored through data analysis using time series, correlation, and regression analysis.
The distribution of numerical columns in the dataset can be seen in Fig. 7. These histograms display the relative distribution of values within a specific interval of the given currency pair or economic indicator. Normally distributed around the current value, currency exchange rates have an unimodal distribution. It has seen a more heterogeneous distribution across economic indicators such as gold, crude oil, and inflation, as well as some unimodal and others bimodal/multimodal histograms. Some histograms also show outliers: data points far from the main distribution. The overall shape and spread of the histograms help to understand small details of the distribution and support, along with further statistical analysis, to identify patterns and trends in the data.
Figure 8 shows the distribution of data points per numerical column using the above table. The sales pile speaks to how many data focuses there are for the given money pair or monetary marker. The plot shows a notable case where the number of data points is very different on columns. Reasons for this variation might include differences in data availability, quality, or nature of the economic indicator. ‘The plot also identifies columns that may have a skewed distribution or too many potential outlier values that potentially need to be explored.’ We created a Bar Plot, which allows a quick visual representation of the data distribution to find patterns and outliers that need attention to a deeper exploratory data analysis.
Visual Comparison of Currency Correlation and Economic Indicators (Fig. 9). Values close to zero indicate no correlation, a positive value indicates a positive correlation, and values near zero indicate a negative correlation. Specific pairs (E.g., AUD/USD & NZD/USD) show robust positive correlations, suggesting they are moving in sync. Economic indicators have a lower correlation with currency pairs, meaning they have a less direct influence on them. If we see outliers in the heatmap, they may be a consequence of specific economic events, human errors, or data anomalies. The heatmap visually represents the correlation matrix, highlighting strong and weak relationships, clusters of correlated variables, and potential outliers that merit further investigation.
As depicted in Fig. 10, the confusion matrix for the LSTM model indicates that it has achieved an overall classification accuracy of 100%. Every instance is ideally classified into its corresponding class, meaning there are neither false positives nor false negatives. There are no errors in the confusion matrix for the model’s prediction of the”Below Threshold”and”Above Threshold”categories, indicating a high level of performance. This means that the LSTM model performs well in capturing the latent distributions and generating the desired future step to achieve accuracy in the problem statement.
Figure 11 Correlation Matrix Heatmap for LSTM Model This correlation heatmap visualizes the correlations between features; each cell in a heatmap represents the correlation coefficient between two variables where positive values mean a positive correlation, negative values mean a negative correlation, and values near zero indicates no correlation. The diagonal elements of the heatmap are equal to 1 since a variable is positively correlated with itself. Some exciting correlations are to be discovered here; I noticed strongly positive correlations between some currency pairs, indicating they move together. However, the correlation between economic indicators and currency pairs is weak. Outlier spikes in the heatmap will be based on financial events and anomalous data entries. So, the heatmap is a valuable tool for visualizing the correlation between different variables in the dataset and can help identify patterns, trends, and potential outliers to investigate further.
The USD/INR exchange rate values between January 1, 2000, and January 12, 2000, for the actual and predicted time series models are illustrated in Fig. 12. A blue line denotes the actual and predicted values through a blue dotted line. We also see from the chart that it follows the trend of the exact values, which means the LSTM model can capture the underlying patterns and make decently accurate predictions. Since the model produced 96% accuracy, we could assume it accurately predicts the results; however, we also notice slight shifts in the predicted values from the original ones, implying that we perhaps need to fine-tune a bit or provide more data.
The actual and predicted value of the USD/INR exchange rate for Episode 2 is shown in Fig. 13. The blue line refers to actual values, and the red dashed line refers to predicted values. The chart shows that the predicted values closely track the trend of the exact values, suggesting that the combination of LSTM and DQN can capture the underlying processes and make relatively accurate predictions. There are, however, some cases in which the forecast is not perfect, indicating that there might be a better fit, either by fine-tuning or fetching further data points. In summary, the image visually illustrates the model’s performance while identifying areas where the model can still be optimized.
Figure 14 is a confusion matrix combining DQN and LSTM, showing 100% classification of all states. Every instance is correctly classified into its class, meaning there are no false positives or negatives. We can see no errors in the confusion matrix, indicating that this model performs very well in classifying the”Below Threshold”and”Above Threshold”categories. This suggests that the joint DQN and LSTM models learned the underlying distributions and could make accurate predictions in this poll.
A correlation coefficient was obtained, and a perfect linear prediction was discovered, as shown in Fig. 15, with a value of 0.9942 for the correlation between the actual and predicted values. Actual values (y-axis) are compared with the expected values (x-axis). The scatter plot is highly concentrated between the actual and predicted values, making it a good indication. This indicates that the integrated LSTM-DQN model fits the massive future and helps predict very well. Such a strong correlation confirms that the model suits this particular job.
After producing a white noise process, we can plot the Autocorrelation Function (ACF) (Fig. 16), which shows the correlations between a time series and its lagged version. Here, the autocorrelation value is close to 0 for most lags and lies within the confidence bands. This means that the time series and its lagged values are almost uncorrelated. In other words, there is no autocorrelation in the time series data – that is, past values do not influence future values significantly. This indicates that the time series is probably stationary, a preferred feature for many time series analysis methods.
Figure 17 compares the performance of three models, MSE, MAE, and RMSE, using LSTM, DQN, and LSTM-DQN. The LSTM model consistently yields better performance for all metrics, which suggests that it makes more accurate and reliable predictions than the other two models. However, the performance of those models depends on factors such as data quality, model complexity, and hyperparameter tuning. An ensemble method could yield an even better classifier than the current best classifier.
The models are CNN, GRU, LSTM, and DQN + LSTM, and performance comparison is provided in terms of mean squared error, mean absolute error, and root mean squared error, as shown in Table 5. Based on these MSE, MAE, and RMSE values, it is evident that, as compared to the FL, DQN, and DQN models, the DQN + LSTM model shows the least values of MSE (0.37), MAE (0.46), and RMSE (0.61), thus indicating its best predictive accuracy and error minimization capabilities. On the other hand, the CNN model shows the highest values for all metrics, as its MSE = 0.86, MAE = 0.96, and RMSE = 0.92, indicating a weak performance. For GRU and LSTM, intermediate results were observed, with LSTM outperforming GRU in terms of the subsequent reduction of standard error.
The comparative analysis of four models (CNN, GRU, LSTM, and DQN + LSTM) is presented in Fig. 18, which utilizes MSE, MAE, and RMSE as metrics. The performance yielded by the DQN + LSTM model is significantly higher than that of other models, with the lowest MSE (0.37), MAE (0.46), and RMSE (0.61). It performs better, implying that the model minimizes errors in the prediction very well and can generalize across unseen data efficiently. Conversely, the CNN model exhibits the lowest performance, with an MSE of 0.86, MAE of 0.96, and RMSE of 0.92, indicating that it needs help to capture the temporal dependencies characteristic of sequential data. The GRU and LSTM models both show average performance, but the LSTM model outperforms the GRU, as evidenced by lower error metrics. The LSTM’s unique architecture comprises memory cells and gating mechanisms that enable it to store critical information across longer sequences. At the same time, GRU employs a more straightforward structure and thus reports the worst output. Its hybrid architecture contributes to the fantastic performance of DQN + LSTM. The LSTM part efficiently handles long-range dependencies, and the Deep Q-Network (DQN) improves action selection via reinforcement learning, whereby the optimal action is learned by observing the results. This enables the model to optimize its predictions, leading to dynamic excellence across metrics. DQN + LSTM is thus an excellent option for sequential data.
Paired t-tests were adopted to test the statistical significance of the forecasting errors (MAE, MSE, RMSE). Results indicated that enhancements provided by the proposed model over baselines were statistically significant (p < 0.05). We will now include a sentence stating this in Section"Experimental results"of the revised manuscript.
We further conduct the ablation study for the combined components of LSTM and DQN, in our case, the combined LSTM-DQN framework, to determine the performance that each factor gives to the overall performance. This study reveals the significance of reinforcement learning and hybrid integration, looking through configuration isolation and analysis, and proves that more robust accuracy and adaptability can be achieved in exchange rate prediction than previous models.
As displayed in Table 6, the ablation study investigates the results of the proposed LSTM-DQN model against the components' versions. Table 4 shows that the full LSTM-DQN model performed the best on all metrics, with the lowest MSE and RMSE of (0.37) and (0.61), respectively, demonstrating the significance of sequential learning and reinforcement-driven optimization synergy. The only LSTM structure achieves high errors due to reduced adaptability, while it can catch some temporal dependencies. The DQN-only model fails to capture temporality and outperforms our proposed temporal modeling unit. Static decision rules or pre-trained DQN layers provide modest gains, yet they cannot match the performance of the hybrid approach. One of the main takeaways from this study is the significance of reinforcement learning to improve prediction and flexibility.
The rest of the sections in this paper are organized as follows: In Section"Related work", we introduce the problem and its motivation and describe the proposed LSTM-DQN algorithm. The performance comparison section compares the LSTM-DQN model to five other state-of-the-art methodologies to highlight accuracy and robustness. Section"Proposed framework"illustrates how the proposed method works and compares its performance with the other five methodologies. Finally, the last section provides the conclusion. Benchmarking is done with the MSE, MAE & RMSE metrics. This finding highlights the effectiveness of the proposed model in capturing temporal dependencies and adapting to dynamic market conditions.
Table 7 compares the performance of the proposed LSTM-DQN model with five advanced models according to the measures of MSE, MAE, and RMSE. The LSTM-DQN has the lowest MSE (0.37) and RMSE (0.61), showing the accuracy and robustness of our prediction. This emphasizes the capability of capturing the underlying patterns regarding exchange rate forecasting.
On the other hand, more importantly, the LSTM-DQN model proposed could reach lower Mean Squared Error (MSE) and Root Mean Square Error (RMSE) due to the ability of the LSTM network to capture long-range temporal dependencies and the reinforcement elements of the DQN which allows the system to learn the best actions to take to achieve the optimal forecast. Thus, LSTM is excellent for sequential data pattern recognition and decreases the appearance of gradient problems so that the model can remember critical long-term dependencies of the exchange rate mutation. The traditional LSTM model learns using historical data only and cannot adapt to sudden changes in the market. By integrating DQN, the model can adjust its forecasting strategy incrementally and learn to make more accurate predictions. Using experience replay allows the model to not overfit on recent trends and generalize well for varying conditions. The target network stabilization technique in DQN ensures stable learning. Otherwise, non-stable weight updates can cause predictive performance to degrade. These improvements enable the suggested framework to adapt dynamically to the evolving patterns in the market, leading to a significant decrease in the level of prediction errors as it outperforms the other benchmark models in terms of decline in the mean square error (MSE) and root mean square error (RMSE).
The strength and flexibility of the LSTM-DQN model can be attributed to several important factors that improve its capacity to process turbulent exchange rate data. LSTM – Long Short-Term Memory (a component of the proposed model) successfully captures long-term dependencies and sequential information from financial time-series data while reducing the issue of vanishing gradient, making it essential for learning long historical trends. Second, the DQN part applies an agile learning system that successively improves the forecasting formulation by using instant feedback signals to modify the timing due to sharp market variations. Third, the model uses experience replay, which avoids overfitting by training on many historical and recent observations, enabling the model to generalize well across different regimes. Fourth, a target network stabilization technique in DQN, which avoids sudden hits, ensures stable learning, which is an asset in noise-destructive domains usually seen in a financial environment with sudden trend reversals. Fifth, this model incorporates macroeconomic indicators like gold and crude oil prices and inflation rates, accommodating external economic forces influencing currency changes. These features make LSTM-DQN highly resistant and flexible for predicting the exchange rate in turbulent markets.
The performance of the proposed model demonstrates its efficiency in making the above combination of LSTM and DQN work for exchange rate forecasting over its counterparts (highly competitive state-of-the-art models). The results for all four parameters are reported in Table 7. Namely, the proposed model shows the lowest Mean Squared Error (MSE) of 0.37, Mean Absolute Error (MAE) of 0.46 and Root Mean Squared Error (RMSE) of 0.61, which at last shows superiority in performance over other benchmark models. This improvement demonstrates the potential of combining LSTM's capability of handling long-term dependencies in time series data with the adaptive optimization feature of the DQN reinforcement learning mechanism to create a more robust and accurate forecasting framework. Deep learning models, including MF-LSTM, AB-LSTM-GRU, and TLS-LSTM, demonstrate robust predictive capabilities but cannot be optimized iteratively by real-time feedback, hence making them less responsive to abrupt market changes. By incorporating DQN, the model learns continuously from past experiences (experience replay) and updates its policy based on reward signals, enabling it to adapt its forecasting strategies. The DQN employs a target network that updates less frequently, maintaining stability in the face of housing pressure behavior in some financial markets, which is not applicable for all cases. It can be concluded from these results that the incorporation of LSTM's ability to find sequential patterns, along with DQN's ability to make adaptive decision-making, greatly enhanced the accuracy, adaptation, and robustness of a model of exchange rate prediction.
Figure 19 provides a comparative analysis of the performance metrics — Mean Squared Error (MSE), Mean Absolute Error (MAE), and Root Mean Squared Error (RMSE) — across five advanced models and the proposed LSTM-DQN model. The proposed model consistently outperforms the other models, achieving the lowest values for MSE (0.37) and RMSE (0.61), indicating superior accuracy in capturing patterns and minimizing prediction errors. Models like Multimodal Fusion-Based LSTM (MF-LSTM) and AB-LSTM-GRU exhibit higher MSE and RMSE values, reflecting their limitations in handling complex and dynamic relationships in exchange rate forecasting. TSMixer, though transformer-based, lacks adaptability to dynamic market fluctuations due to the absence of reinforcement learning mechanisms, resulting in moderate performance. The critical success of the proposed LSTM-DQN model lies in its hybrid approach. By integrating the temporal dependency learning capability of LSTM with the adaptive decision-making power of the DQN agent, the model iteratively improves its predictions using feedback-driven optimization. This reinforcement-driven adaptability allows the model to adjust to dynamic market conditions and capture non-linear relationships effectively. Overall, the proposed LSTM-DQN model demonstrates its ability to address the limitations of existing models, providing accurate, robust, and adaptive forecasting capabilities essential for the volatile nature of financial markets.
Key insights regarding the effectiveness of the proposed LSTM-DQN model are derived from the experimental results on USD/INR exchange rate data. Firstly, the model achieves the lowest MSE (0.37), MAE (0.46), and RMSE (0.61), outperforming state-of-the-art deep learning methods, indicating its superior predictive ability. It emphasizes the merit of combining LSTM learning long-term temporal dependencies with the DQN method to optimize adaptively so that the model can continuously adjust predictions based on updates in market dynamics. Second, traditional models based on LSTM, e.g., MF-LSTM and TLS-LSTM, can learn sequential patterns effectively. Still, they lack an adaptive learning mechanism, which results in limited performance in volatile market conditions. Thirdly, due to the real-time decision-making process in reinforcement learning execution, predictive performance can be improved as experience avoids overfitting and aids in model generalization across distinct market environments. Fourth, and in addition to the macroeconomic indicators used in the model (gold prices, crude oil prices, inflation rates), ensuring that the model maximizes predictive power may also help by accounting for economic factors contributing to currency changes. Finally, results testify that hybrid deep learning and reinforcement learning outperform purely supervised learning in time series prediction for the financial domain. The findings indicate how AI-based adaptable forecasting frameworks can empower financial analysts, policymakers, and forex traders to manage risks better and make well-informed investment choices.
Results indicate the performance of LSTM models in terms of MSE and RMSE. An MSE of 0.37 and an RMSE of 0.61 represent a significant performance for exchange rate forecasting, primarily due to the notable decrease in the prediction error compared to previous models. Foreign exchange rate markets are notoriously unstable, and even small gains in prediction quality may lead to significant progress in financial management, risk prevention, and trading operations. A lower MSE and RMSE suggest that the model could reduce large deviations between the predicted and actual exchange rates and produce more accurate and stable forecasts. In finance, the importance of minimizing prediction error margins is paramount; for instance, although inaccurate predictions on currency price movements could lead to economic losses for forex traders and institutional investors, inaccurate predictions on price movement trends could jeopardize currency policies implemented by decision-makers. Most importantly, when compared to traditional models such as MF-LSTM (MSE: 0.62, RMSE: 0.79) and TSMixer (MSE: 0.49, RMSE: 0.70), the error metrics of the LSTM-DQN model are lower, reinforcing the notion that reinforcement learning significantly increases a model's adaptability, and thus more accurately captures sudden shifts in the market. So, a very low RMSE value of 0.61 further guarantees that the model's forecasts likely remain consistent and robust in various market conditions; thus, it is usable in practical settings in real-world financial settings. The significance of these results lies in proving deep reinforcement learning as an effective method of merging deep learning with reinforcement learning to reduce forecasting uncertainty, which translates into tactical superiority in financial decision-making and risk reduction mechanisms.
This table compares MSE, MAE, and RMSE metrics for each model. It highlights the significant performance improvements the proposed LSTM-DQN approach provides compared with existing state-of-the-art methods. The proposed model achieves the lowest MSE (0.37), MAE (0.46), and RMSE (0.61) compared to 67 other models, including MF-LSTM (MSE: 0.62, RMSE: 0.79), AB-LSTM-GRU (MSE: 0.54, RMSE: 0.73), and TSMixer (MSE: 0.49, RMSE: 0.70), as seen in Table 7. Though hybrid deep learning models can also outperform traditional LSTM variants (for example, TLS-LSTM: MSE: 0.44, RMSE: 666), such models cannot match the proposed model either, as they lack adaptive optimization mechanisms. The use of reinforcement learning (DQN) with LSTM-DQN allows for continuous updates of the model based on online feedback, which helps the model adapt to stock market changes over time, and this can be considered an advantage of the LSTM-DQN model compared to traditional deep learning models. Moreover, due to experience replay and target networks, DQN is more stable and better at generalizing, reducing the risk of overfitting historical patterns. Combining the Sequential Learning ability of LSTM and the Reinforcement-based optimization of traditional DQN leads to highly predictive accuracy and a robust model for financial forecasting vs. a purely supervised approach, since the statement result.
The numerous essential findings derived from the experimental results further substantiate the effectiveness of the proposed LSTM-DQN framework in reducing forecast errors. In terms of results, the LSTM-DQN model induced the lowest MSE (0.37) and RMSE (0.61) with significantly improved performance against benchmark models such as MF-LSTM (MSE: 0.62 RMSE: 0.79), AB-LSTM-GRU (MSE: 0.54, RMSE: 0.73), and TSMixer (MSE: 0.49, RMSE: 0.70). The effectiveness of the proposed framework is demonstrated by the lower error values, indicative of its higher accuracy and stability compared to traditional models of deep learning. Second, by applying DQN, this model also enables reinforcement learning, which means it can formulate real-time optimization to generate differential model outputs according to market information. This is especially important in financial forecasting, where static models are inadequate for large-scale, sudden changes in the market and volatility. DQN is also robust due to its experience replay mechanism, which ensures you use experience and that the model doesn't drift too much with a few data points in time. In addition, target network stabilization promotes smoother learning and limits sudden movements of predictions, making the model more reliable. By leveraging the sequential pattern recognition capability of LSTM and the optimized decision-making characteristic of DQN, the proposed framework exhibits excellent transferability to various market environments, leading to its superior predictability for exchange rates with the least forecast error.
The proposed LSTM-DQN model exhibits high adaptability to real-time fluctuating market conditions, making it a suitable candidate for dynamic exchange rate forecasting. Traditional deep learning models like MF-LSTM, AB-LSTM-GRU, and TLS-LSTM work on historical patterns, whereas in addition to historical data, we use reinforcement learning via Deep Q-Networks (DQN), allowing the model to adapt predictions based on general changes in market trends over time. This is accomplished using real-time feedback systems, such that the model constantly updates its forecasting policy based on new data. As such, this allows the model to generalize better, as it balances between too short a time frame (where it would just fit some small-scale noise) and too long a time frame (where it would end up following the market around downwards to avoid short-term fluctuations). The DQN training algorithm also employs a target network stabilization technique that controls learning updates to the network weights, helping prevent wild weight fluctuations and enabling the model to be robust to high volatility and sudden drops in price during the market. As revealed in Table 7, the empirical outcomes demonstrate how the LSTM-DQN model surpasses traditional deep learning models in yet again variable financial conditions, supporting its effectiveness for reconfiguring its response to erratic signals. This additional flexibility is essential for applications in the broader financial markets because it allows more accurate, responsive, and risk-sensitive exchange rate predictions, which are critical for decision-making among traders, analysts, and policymakers in rapidly changing global markets.
Table 8 demonstrates a qualitative comparison of the proposed LSTM-DQN model against five cutting-edge models, namely, MF-LSTM, AB-LSTM-GRU, TSMixer, Hybrid LSTM Models, and TLS-LSTM; it has been evaluated on prominent features including adaptability, temporal dependence articulation, reinforcement learning, computational overhead, scalability, precision, interpretability, and incorporation of decision-making. ii) The LSTM-DQN outperforms the other models by being a reinforcement learning model that optimizes the algorithm iteratively over each input and leverages the market's unique characteristics when highly dynamic. On the other hand, models such as MF-LSTM and AB-LSTM-GRU are based on an offline strategy (static strategy in AB-LSTM-GRU and ensemble strategy in MF-LSTM) with no reinforcement learning included, which in action responds to the environment. LSTM-DQN is unique in its integration and, therefore, optimization of the decision-making process based on the predicted data and feedback, as all other models rely on temporal solid dependency modeling. The LSTM-DQN model expends more computational complexity resources using a reinforcement learning mechanism. Its scalability is still considered high since it employs GPU acceleration to process large datasets efficiently. The proposed model also excels in accuracy, showing superior performance in quantitative metrics such as MSE and RMSE compared to other models. Interpretability is moderate since reinforcement learning can be pretty complex, but features that integrate prediction and decision-making, which the other models lack, are essential here. The results from this study suggest that the proposed LSTM-DQN model is a robust and effective tool for exchange rate prediction, making it a valuable addition to the range of models available for exchange rate forecasting in turbulent markets.
Given their intrinsic design constraints, other models might fail to execute iterative optimization based on real-time feedback. Traditional statistical methods, such as ARIMA, are inherently grounded on predefined mathematical assumptions and are inherently static, meaning they cannot adapt to changing market conditions, rendering them unsuitable for real-time decision-making. Deep learning techniques like LSTM, GRU, and CNN-based architectures work on a supervised learning model, meaning they train via historical data but do not fine-tune their predictions after deployment. These models need to be retrained on a new dataset to enable learning of the latest processing patterns, which can be costly and time-consuming. Additionally, deep learning models do not have an explicit learning mechanism based on reward; hence, they cannot learn by interacting with the market environment and, thus, continually improve. What they lack (unlike reinforcement learning approaches) is an iterative process to ameliorate their decision-making via exploration-exploitation strategies. Moreover, conventional deep learning models are also prone to data drift — if there are significant shifts in market conditions, the model underperforms since it relies on outdated patterns. In contrast, SSE-DQN alleviates these difficulties through Deep Q-Learning, which adapts and updates its forecasting model based on the Q-value at each step. It has stronger adaptability and robustness in volatile financial markets.
This study offers valuable insights for financial decision-making, including insights from forex traders, risk managers, and monetary policymakers. Thus, the proposed framework combining deep learning with reinforcement learning could improve exchange rate forecasting by achieving more accuracy and adaptivity in the data. It would help traders and institutional investors make more informed decisions to optimize their trading strategy and reduce losses in unstable foreign exchange markets. By incorporating vital macroeconomic factors, this model boasts a broader context to understand the currency better and help the financial analyst make better market predictions. Furthermore, policymakers and central banks could also benefit from these improved forecasting capabilities to formulate effective policies involving inflation management and financial market stabilization. It also emphasizes the transformative use of AI-driven decision-making in financial applications, indicating future scope for research on hybrid models that offer improved predictive accuracy and real-time responsiveness in their complexities.
