Submit to this Journal Review for this Journal Propose a Special Issue

Article Menu

Share Help Cite Discuss in SciProfiles

Open AccessArticle

Peer-Review Record

Hybrid Deep Learning Techniques Integrated with Machine Learning for Foreign Exchange Rate Forecasting

Electronics 2026, 15(7), 1463; https://doi.org/10.3390/electronics15071463

by Yu Cui^*

and Jingjing Jiang

Reviewer 1: Anonymous

Reviewer 2:

A. Datta

Reviewer 3: Anonymous

Electronics 2026, 15(7), 1463; https://doi.org/10.3390/electronics15071463

Submission received: 15 January 2026 / Revised: 9 March 2026 / Accepted: 16 March 2026 / Published: 1 April 2026

(This article belongs to the Special Issue Celebrating the 70th Anniversary of Beijing University of Posts and Telecommunications—Computer Science and Engineering)

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

Please refer to the attachment.

Comments for author File: Comments.pdf

Comments on the Quality of English Language

Please refer to the attachment.

Author Response

Review Report 1

Manuscript: Electronics-4126259

Recommendation: Major Revision

Overall assessment. The manuscript contains multiple objective, verifiable inconsistencies (task definition, notation, metrics, and figure references) that directly undermine technical validity and reproducibility. I recommend major revision.

Major issues (must be addressed).

Task-definition mismatch (Forecasting vs Detection). The title states “Forecasting”, but keywords and the introduction repeatedly use “detection” (e.g., “Foreign exchange rate detection”, “Forex detection remains...”, “detection task/systems”). The task must be made consistent throughout.

Response: The foreign exchange rate forecasting term is verified throughout the document based on given suggestion.

Iterative Imputer: definition vs pseudocode contradiction. The text defines iterative imputation as feature-wise regression using other features iteratively, but the pseudocode applies imputation per-sample (e.g., x′ i = IterativeImputation(xi)). Align the algorithm description and specify the exact fit/transform protocol (including leakage control).

Response: The definition of input variables is revised in algorithm as we defined x in earlier. The updation in algorithm in line 487 “Algorithm: Exchange rate forecasting algorithm using BiLSTM+Attention+RFR

Input:

Output:

Data:

Params:

For each

Notation error + undefined symbol. After defining H = [h1, . . . , hT ], the manuscript states “hT ... at time t” (indexing is inconsistent; should be ht). The same context introduces hs (or similar) without definition. Fix indices and define all symbols.

Response: The terms are changed based on given suggestion in line 356-359“In this architecture, BiLSTM has a dual role. First, it serves as a feature extractor and not as a predictor. The BiLSTM gains enriched representations of the sequences by generating representations from past and future contexts in a bidirectional fashion. The encoded temporal features are defined as where displays the hidden state at time t and provides context-aware embeddings to capture the long-term dependencies within the time series.

Undefined input symbol in pseudocode (x = RT ). The pseudocode uses x = RT , but R is not established earlier (meaning and dimensions are missing). Define R (and its shape) or use the already-defined data notation consistently.

Response: We already defined input as “Input: ” so now I replaced with previous definition. Changes made in line 422

Metric contradiction (Pearson vs Spearman). Table 3 is labelled Pearson Correlation, while the main text reports Spearman correlations (e.g., 0.924, 0.989). The manuscript must use one metric consistently (and recompute/relable accordingly).

Response: the metrics are checked and it is Pearson correlation, it is misspelled with confusion. Now the computation is checked and text is replaced. Thank you so much for correcting my mistakes with patients.

Minor issues (presentation/formatting; still must be fixed).

Language polishing. Multiple sentences show broken phrasing/word choice that obscures meaning (e.g., “halting” used for attention; fragmented sentences in the method description). A thorough copy-edit is needed.

Recommendation. Given the above major issues affecting correctness, internal consistency, and reproducibility, I recommend Major Revision.

Response: Language polishing is done based on given suggestion

Reviewer 2 Report

Comments and Suggestions for Authors

The authors have presented very effectively their novel hybrid deep learning model for forecasting foreign exchange rates. Overall the manuscript is well organized but there needs to be couple of aspects to be taken into account to make the manuscript more effective and content specific. Their algorithm is good but their literature review and comparison models needs significant review. Below are areas to be explained or taken care of.

It is a bit surprising to see in the literature survey section; the authors didn’t mention previous research in Forex prediction models. Most of their literature review mentioned forecasting models for data related to solar irradiance, stock market, surface settlement, wave vehicular communication and wind speed. All of them are non-Forex data sets and their dynamics and behavior are quite different from Forex data. There are several research models handling Forex data forecasting. It would be good to give their reference. I have listed some of the recent manuscripts among many below for their reference which talk about models being used for exchange rate predictions.
1. https://www.mdpi.com/2227-9091/12/9/139
2. https://www.sciencedirect.com/science/article/pii/S2772662223000334
3. https://ejournal.uin-suska.ac.id/index.php/IJAIDM/article/view/38354
4. https://www.sciencedirect.com/science/article/pii/S246822762500287X

For comparison of their method with other methods instead of comparing with simpler methods, they should have compared with existing robust methods which have been used in the past to predict Forex data like MVO-BiGRU, MVO-BiLSTM, TCN-LSTM etc. which have been referred to in the above publications.

I am curious to know why the authors removed non-numeric data from the dataset. Removing country name or code makes sense but if the data contains other non-numeric data, they should keep them since BiLSTMs can handle non-numeric data after converting them into numerical tensors. It would be good if they mention about the dataset a little more for the reader and convert relevant non-numeric columns into tensors and feed to the model.

For timeseries Forex data, with missing data, efficient imputing algorithms like BRITS or Kalman Smoothing should be used because they capture both temporal and cross-sectional data dependencies. I am curious if the authors have thread these methods before trying iterative imputation.

To stabilize variance in Forex time series data, power transformations like Box-Cox or Yeo-Johnson transformations is work much better than Log transformation which the authors have used.

For timeseries data feature scaling methods, RobustScaler is often superior to MinMaxScaler (used by the authors) or StandardScaler because financial data is prone to significant outliers and volatility. RobustScaler handles these outliers more effectively.

Did the authors perform any feature engineering? Or was it not required at all?

In the methods section, the authors have given too much details of the data transformation method (log-transformation), data imputation method (iterative imputer), and feature scaling method (MinMaxScaler). All of these methods are very establised methods and don’t need to be again explained in the manuscript there by making it long.

English lines or word corrections:

Line 187 ends abruptly.
What is uhistorical in line 216?
What is datenderstanding in line 217? Is it understanding instead?

Comments for author File: Comments.pdf

Author Response

Reviewer comments-2

1) It is a bit surprising to see in the literature survey section; the authors didn’t mention previous research in Forex prediction models. Most of their literature review mentioned forecasting models for data related to solar irradiance, stock market, surface settlement, wave vehicular communication and wind speed. All of them are non-Forex data sets and their dynamics and behavior are quite different from Forex data. There are several research models handling Forex data forecasting. It would be good to give their reference. I have listed some of the recent manuscripts among many below for their reference which talk about models being used for exchange rate predictions.

https://www.mdpi.com/2227-9091/12/9/139
https://www.sciencedirect.com/science/article/pii/S2772662223000334
https://ejournal.uin-suska.ac.id/index.php/IJAIDM/article/view/38354
https://www.sciencedirect.com/science/article/pii/S246822762500287X

Response: Section 3 and reference sections are fully rewritten based on given suggestion. The irrelevant articles are removed from various forecasting; now full literature is about foreign exchange forecasting.

2) For comparison of their method with other methods instead of comparing with simpler methods, they should have compared with existing robust methods which have been used in the past to predict Forex data like MVO-BiGRU, MVO-BiLSTM, TCN-LSTM etc. which have been referred to in the above publications.

Response: comparison of the SOTA models are DISCUSSED IN TABLE 4

Table 4. Comparative analysis of state of art models with proposed

Model	R²	MAE
MVO-BiGRU (Saudi-euro) [25]	0.984	0.021
MVO-BiGRU(china-euro) [25]	0.962	0.006
TCN-QV [26](US-China)	0.997	1.556
SARIMA-LSTM (Africa vs US,EUR,CYN) [27]	-	0.006
BiLSTM + Attention + RFR	0.985	0.118

The results from the table 4 shows recent models performed well, yet the R2 and MAE statistics reveal differences among them. Each of the hybrid deep learning models demonstrated strong forecasting capability, but with varying degrees of accuracy and error levels. The best performing model was the MVO-BiGRU model, demonstrating high performance across both the Saudi-Euro and China-Euro currency pairs with an R2 of .984 and .962 respectively, and with notably low MAE values, indicating that it is robust across a variety of currency pairs. Although the TCN-QV model exhibited the largest R2 at .997 for the US-China pair, this model also exhibited the largest MAE at 1.556, thus limiting its potential for reliable use. Lastly, the SARIMA-LSTM model demonstrated a small MAE (.006) when comparing forecasts of African currencies to those of the US, EUR and CYN, but did not report an R2 statistic, which limits the ability to fully assess its relative performance compared to the other models examined. While the BiLSTM + Attention + RFR model reported a large R2 (.985), the MAE (.118) was significantly larger, indicating that while the model has good explanatory power, the precision of its forecasts is likely lower than many of the hybrid models used in comparison. The existing models all for certain country specific were there is no real time dynamics, where proposed model is global foreign exchange forecasting were all countries are data are utilized to train the model.

3) I am curious to know why the authors removed non-numeric data from the dataset. Removing country name or code makes sense but if the data contains other non-numeric data, they should keep them since BiLSTMs can handle non-numeric data after converting them into numerical tensors. It would be good if they mention about the dataset a little more for the reader and convert relevant non-numeric columns into tensors and feed to the model.

Response: The dataset used in this study contains only two non-numeric columns: Country Name and Country Code. These fields serve purely as identifiers and do not contain intrinsic temporal or economic information relevant to exchange-rate forecasting. Apart from these identifiers, all remaining columns correspond to annual numerical exchange-rate values.

We agree that Bidirectional LSTM networks are capable of incorporating categorical variables once they are converted into numerical tensors (e.g., via one-hot encoding or embeddings). However, in this case, the non-numeric columns do not represent explanatory variables but rather static identifiers. Encoding country identifiers (e.g., through one-hot encoding) would introduce fixed country labels instead of dynamic predictive features and would substantially increase input dimensionality without adding new economic information. Since the objective of this work is to model temporal dynamics using historical exchange-rate sequences, the model was trained exclusively on numerical time-series values.

To improve model efficiency we have done one hot encoding and proposed data preprocessing is updated in section 3.1

4) For timeseries Forex data, with missing data, efficient imputing algorithms like BRITS or Kalman Smoothing should be used because they capture both temporal and cross-sectional data dependencies. I am curious if the authors have thread these methods before trying iterative imputation.

Response: We acknowledge the reviewer’s comments regarding the use of BRITS and Kalman Smoothing for time-series imputation. Such methods are indeed useful for datasets that have high-frequent observations, and for this reason, we believe BRITS and Kalman Smoothing are not suited for our research. This is because our dataset only has annual observations of exchange rates over several decades. In our dataset, there also seem to be completely random gaps in the observations of exchange rates, rather than with gaps that are long and are a contiguous series of observations.

For this reason, we used Iterative Imputation, which is able to efficiently capture dependencies between cross features by country, by year, and do so within a reasonable time in a global dataset that is large. BRITS and Kalman approaches are less suited to global datasets because they have macroeconomic datasets that are long and annual. We appreciate the reviewer’s suggestions, and we will consider advances in the imputation methods in datasets with exchange rates that are of a higher frequency in the future.

5) To stabilize variance in Forex time series data, power transformations like Box-Cox or Yeo Johnson transformations is work much better than Log transformation which the authors have used.

Response: We appreciate the reviewer’s comments on possible other techniques for variance-stabilization. Indeed, power transformations like Box-Cox and Yeo-Johnson can be effective for variance normalization, but given the presence of zeros and low values (for example, pegged or fixed-rate regimes in earlier years) in the foreign exchange dataset, the applicability of Box-Cox is limited without additional data shifting.

We chose log1p for the logarithmic transformation because it:

Handles zero-valued observations naturally,
Adjusts right-skewness which is often present in exchange-rate distributions,
Retains interpretability in the context of finance, and
Provided reliable empirical results in the preliminary assessment.

However, the manuscript has been modified to recognize Box-Cox and Yeo-Johnson transformations as other viable options, which may be considered in future extensions of this research.

6) For timeseries data feature scaling methods, RobustScaler is often superior to MinMaxScaler (used by the authors) or StandardScaler because financial data is prone to significant outliers and volatility. RobustScaler handles these outliers more effectively.

Response: We thank the reviewer for the suggestion about the RobustScaler. We agree that financial time-series data are particularly subject to outliers and volatility. For this study, extreme values were log transformed and imputed iteratively (i.e. extreme values were kept to reduce the influence of heavy-tailed distributions).

Min-Max scaling was used because it keeps the relative temporal dynamics within a fixed range, which is beneficial for LSTM (Long Short Term Memory) based technologies that are sensitive to the input scale. In this situation, We found the performance of Min-Max and RobustScaler to be similar, with Min-Max achieving faster training. Still, the manuscript discusses RobustScaler in greater detail as a viable option for data sets with more extreme outliers.

7) Did the authors perform any feature engineering? Or was it not required at all?

Response: Rather than relying on manually engineered financial indicators, this study adopts an implicit feature engineering approach. Temporal sequencing of exchange-rate values enables the BiLSTM network to learn lag dependencies directly from historical data. Variance stabilization through log transformation, iterative imputation of missing values, and Min–Max normalization further reshape the feature space to enhance learning stability. Additionally, the attention mechanism and Random Forest Regressor perform automatic feature selection and nonlinear interaction learning, eliminating the need for handcrafted features while improving robustness and generalization.

8) In the methods section, the authors have given too much details of the data transformation method (log-transformation), data imputation method (iterative imputer), and feature scaling method (MinMaxScaler). All of these methods are very establised methods and don’t need to be again explained in the manuscript there by making it long.

Response: Based on suggestion I have deleted less important para in the manuscript.

9) English lines or word corrections:

Line 187 ends abruptly.
What is uhistorical in line 216?
What is datenderstanding in line 217? Is it understanding instead?

Response: All the lines are checked and r5espective changes are made based on suggestion.

Reviewer 3 Report

Comments and Suggestions for Authors

This paper presents a hybrid forecasting framework that combines a Bidirectional LSTM, an attention mechanism, and a Random Forest Regressor to predict annual foreign exchange rates using World Bank data for approximately 200 countries over the period 1960–2024. The proposed architecture is technically well motivated and the experimental section is extensive. However, several aspects of the research design, evaluation methodology, and presentation limit the strength of the contribution and should be addressed before the paper can be considered for publication.

A primary concern relates to the evaluation strategy adopted for what is fundamentally a time-series forecasting problem. The authors rely mainly on an 80/20 random train–test split, which is not appropriate in this context, as it risks temporal leakage by allowing future observations to influence model training. Although rolling-origin (walk-forward) validation is mentioned later in the paper, it appears as a supplementary robustness check rather than the central evaluation framework. For forecasting applications, walk-forward validation should form the core of the experimental design to ensure that reported performance reflects realistic, forward-looking predictive ability.

The strength of the naïve baselines further raises questions about the practical significance of the proposed hybrid model. In particular, the “Last Observed Value” baseline achieves an R² of 0.891, indicating strong year-to-year persistence in exchange rates at the annual frequency. While the proposed model improves this to an R² of 0.985, the incremental gain over such a simple persistence model warrants more careful justification in terms of real-world usefulness. Moreover, the model relies exclusively on past exchange rate values and does not incorporate macroeconomic fundamentals such as interest rate differentials, inflation, GDP growth, trade balances, or monetary policy indicators, which are well-established drivers of exchange rate dynamics. This omission limits both the economic interpretability and the practical applicability of the approach. At a minimum, this limitation should be discussed more explicitly, or framed as a key direction for future work.

There are also issues of internal consistency in the reported results. Specifically, the performance metrics shown in Table 3 do not match those reported in Figure 15. Table 3 reports MSE = 0.080 and MAE = 0.118 for the proposed model, whereas Figure 15 shows higher error values (MSE = 0.102 and MAE = 0.148). These discrepancies are not explained and undermine confidence in the reported performance unless clarified.

With respect to related work, the literature review covers a broad range of application domains, including solar irradiance forecasting, water quality prediction, and stock price modeling. However, it contains relatively few references that are directly relevant to foreign exchange or currency forecasting using deep learning methods. As a result, it is difficult to clearly position the proposed contribution within the existing body of work on currency and financial time-series prediction. Recent studies on LSTM-based and hybrid models for forex and cryptocurrency markets (e.g., but not limited to, https://doi.org/10.1016/j.mlwa.2025.100648, https://doi.org/10.1016/j.eswa.2024.125457 or https://doi.org/10.1007/978-3-031-74633-8_45) would provide a more appropriate comparative context and help clarify how the proposed approach advances the state of the art in this specific domain.

Reproducibility is another concern. Key architectural and training details are missing, including the number of BiLSTM layers, hidden units, dropout rates, and the main Random Forest hyperparameters such as the number of trees, maximum depth, and minimum samples per leaf. Without these details, it is difficult for other researchers to replicate or fairly compare against the proposed model.

Finally, there are several minor but non-negligible presentation issues. The LIME analysis refers to features labeled “Feature 63,” “Feature 62,” and so on, without explaining what these features represent. Since the input appears to consist of historical yearly exchange rates, this mapping should be made explicit and interpreted in economically meaningful terms. There is also an inconsistency between the figure caption and the main text regarding whether the LIME example corresponds to “Sample 16” or “Sample 18.” In addition, the model name is written inconsistently throughout the manuscript (e.g., “BiLSTM” versus “Bi-LSTM”), which should be standardized.

Comments on the Quality of English Language

Line 216: "uhistorical and up to datenderstanding" appears to be an editing artifact
Line 456: References "Spearman's correlation" but the paper uses Pearson correlation

Author Response

Reviewer 3

Response: The contributions are added in the introduction section line number 83 to 105 “These research contributions are

1.We develop a new hybrid framework that combines Bidirectional LSTM (BiLSTM), additive attention, and Random Forest Regression (RFR). In contrast to most deep learning models, which focus on neural outputs, proposed framework utilizes BiLSTM-attention for temporal feature extraction, while RFR for non-linear ensemble feature refinement, thereby increasing the robustness and generalization of the model for long-horizon exchange-rate forecasting.This research focused on single model frameworks, which incorporate deep sequential and machine learning fusion for improved performance.

2.This research encodes next-year exchange-rate forecasting as a supervised learning problem in a cross-country context. The model leverages each country’s long-term annual historical sequence (1960–2023) to predict the country’s value for the next year.

Such a configuration manifests the model’s learning of structural parallels and nonlinear temporal intricacies across about 200 countries. This a global scale rarely captured in annual foreign exchange forecasting.

4.The following sections in manuscripts are organized as follows. Section 2 shows literature survey followed by section 3 discussion on clear methodology. The results are presented in section 4 and section discusses the conclusion with future scope.

Response: The split ration is corrected and model is rerun with one hot encoding and correct split without leakage . Respective figures also included in result section. The reworked model doesn’t show any major differences in predction regression and error rate. Only pont wise differences we found.

Response: 4.1 Limitations

The hybrid BiLSTM + Attention + Random Forest framework shows predictive capabilities, but there are limitations. The naive persistence baseline has a high R-squared value (R2=0.891), which shows that there is a strong temporal autocorrelation present with the annual exchange-rate series. The proposed model does improve that value to R2=0.985, but the additional gain needs to be evaluated because it can be expected that exchange rates will change slowly over the years which makes it very difficult to beat the naive persistence. The proposed model also has a significant amount of improvement with much lower prediction errors (lower MSE and MAE) which demonstrates an increase in practical value where a small forecast change can improve financial value.

Another limitation involves the model used in the present study because it is based solely on historical exchange-rate data. While this model is designed to remove other macroeconomic fundamentals, it also prevents the model from incorporating elements such as interest-rate differentials, inflation, GDP growth, trade balances, and monetary policy which etc. can all drive exchange-rate changes. By not including these elements, the prediction remains less interpretable. The third aspect concerns our dataset’s annual observations, which smooth out the effects of short-term volatility and structural breaks. While this assists with forecasting over the long-term, high-frequency dynamics, which are of interest to traders and policy analysts, may be obscured.

Subsequent to this, main objective is built on the framework proposed (herein) by including macroeconomic and financial variables, as well as higher-frequency and cross-sectional data from multiple sources. The addition of such exogenous variables is expected to improve the interpretive power, responsiveness to structural breaks, and the forecasting system's usability in practice.

Response: Figures are corrected after model is retrained with corrected values. The values are updated and figures are corrected.

https://doi.org/10.1016/j.mlwa.2025.100648, https://doi.org/10.1016/j.eswa.2024.125457 or https://doi.org/10.1007/978-3-031-74633-8_45) would provide a more appropriate comparative context and help clarify how the proposed approach advances the state of the art in this specific domain.

Response: The literature fully revised based on given suggestion. Above relevant article is included and studied and presented in sota table 5.

Table 5. Comparative analysis of state of art models with proposed

Model	R²	MAE
MVO-BiGRU (Saudi-euro) [25]	0.984	0.021
MVO-BiGRU(china-euro) [25]	0.962	0.006
TCN-QV [26](US-China)	0.997	1.556
SARIMA-LSTM (Africa vs US,EUR,CYN) [27]	-	0.006
TCN+LSTM [28]	-	0.240
BiLSTM + Attention + RFR	0.985	0.118

The results from the table 5 shows recent models performed well, yet the R2 and MAE statistics reveal differences among them. Each of the hybrid deep learning models demonstrated strong forecasting capability, but with varying degrees of accuracy and error levels. The best performing model was the MVO-BiGRU model, demonstrating high performance across both the Saudi-Euro and China-Euro currency pairs with an R2 of .984 and .962 respectively, and with notably low MAE values, indicating that it is robust across a variety of currency pairs. Although the TCN-QV model exhibited the largest R2 at .997 for the US-China pair, this model also exhibited the largest MAE at 1.556, thus limiting its potential for reliable use. Lastly, the SARIMA-LSTM model demonstrated a small MAE (.006) when comparing forecasts of African currencies to those of the US, EUR and CYN, but did not report an R2 statistic, which limits the ability to fully assess its relative performance compared to the other models examined. While the BiLSTM + Attention + RFR model reported a large R2 (.985), the MAE (.118) was significantly larger, indicating that while the model has good explanatory power, the precision of its forecasts is likely lower than many of the hybrid models used in comparison. The existing models all for certain country specific where there is no real time dynamics, where proposed model is global foreign exchange forecasting were all countries are data are utilized to train the model.

Response: the architecture summary is added as table in section 3 for reproducibility. Thank you so much for pointing minute mistakes to improve this article.

Response: The report added in line 739 “In the LIME visualization, the attributes designated as “Feature k” are the historical yearly exchange-rate inputs. More specifically, Feature 63 is 2023, Feature 62 is 2022, Feature 61 is 2021, and so on, in chronological order from oldest to most recent. Therefore, the explanation shows that the most recent historical exchange-rate values are the most influential to the prediction made by the model. This supports the theory in financial time-series that recent market states are more valuable than older historical data for predictive purposes.

Comments on the Quality of English Language

Line 216: "uhistorical and up to datenderstanding" appears to be an editing artifact
Line 456: References "Spearman's correlation" but the paper uses Pearson correlation

Response: These lines are corrected in the article based on given suggestion.

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

I am satisfied with the author's revisions and suggest accepting this manuscript.

Author Response

Thank you for your valuable comments.

Reviewer 2 Report

Comments and Suggestions for Authors

The authors did take care of the suggestions.

Author Response

Thank you for your valuable suggestions to improve the manuscript.

Reviewer 3 Report

Comments and Suggestions for Authors

The authors have made substantial efforts to address the concerns raised in the first round of review. The evaluation strategy has been corrected to use a proper chronological train/validation/test split, the literature review now includes relevant forex-specific references, key hyperparameters have been documented, and the LIME feature interpretation has been clarified. These revisions meaningfully strengthen the paper.

However, several issues remain that require attention before the manuscript can be accepted. These are detailed below.

1) In the original submission, I noted a discrepancy between the performance metrics reported in Table 3 and those shown in Figure 15. The revised Table 3 now reports MSE = 0.078 and MAE = 0.108 for the proposed model, while Figure 15 has been replaced with country-specific time-series plots (UK and India) that no longer display aggregate metrics.

While the discrepancy no longer appears in the current version, it remains unclear whether the original inconsistency was due to an error or to metrics being computed on different data partitions. To ensure transparency and reproducibility, please clarify:

Are the metrics in Table 3 computed exclusively on the 2024 test set?
What was the source of the discrepancy in the original submission?

2) The Limitations section (lines 626–632) now acknowledges that the naïve persistence baseline achieves R^2 = 0.887 due to strong temporal autocorrelation in annual exchange rates. While this is an improvement over the original submission, the justification for the hybrid model's practical value remains somewhat limited.

The authors argue that lower MSE and MAE values demonstrate practical improvement. However, given that annual exchange rates are highly persistent by nature, a more convincing demonstration would include:

A discussion of specific forecasting scenarios where the hybrid model's incremental accuracy gains translate to meaningful economic or decision-making value
Alternatively, an analysis of performance during periods of structural breaks or high volatility, where the naïve baseline would be expected to fail

The added statistical significance test (Table 4) compares the hybrid model against Random Forest alone, which is not the most policy-relevant comparison. A paired test against the naïve persistence baseline would more directly address whether the complexity of the hybrid model is justified.

3) Table 2 now reports several key hyperparameters (BiLSTM layers, units, dropout, number of RF trees), which is a significant improvement. However, the Random Forest hyperparameters remain incomplete. The text mentions that Grid Search was used to tune parameters including maximum tree depth (d_max) and minimum samples per leaf (s_min) (lines 344–346), but the selected values are not reported.

Please add the final hyperparameter values selected by Grid Search to Table 2 or the main text to ensure full reproducibility.

4) The revised manuscript still contains numerous grammatical errors and unclear phrasing. Examples include:

Lines 463–464: Inconsistent use of "Bi-LSTM" vs. "BiLSTM"
Lines 623–624: The sentence beginning "The existing models all for certain country specific..." is grammatically unclear

A thorough proofreading pass is strongly recommended before final publication.

Comments on the Quality of English Language

refer topoint 4 of comments

Author Response

Reviewer 3

However, several issues remain that require attention before the manuscript can be accepted. These are detailed below.

In the original submission, I noted a discrepancy between the performance metrics reported in Table 3 and those shown in Figure 15. The revised Table 3 now reports MSE = 0.078 and MAE = 0.108 for the proposed model, while Figure 15 has been replaced with country-specific time-series plots (UK and India) that no longer display aggregate metrics.

Response: Table 3 shows the avg of model performance by combining all countries in the dataset. For the clarity data is added below figure 15 as “Figure 15 shows the actual and predicted data on sample of two countries for the visual understanding. The prediction curve shows that each country actual data is nearby to predicted data. The overall error values are reported in table 3.

Are the metrics in Table 3 computed exclusively on the 2024 test set?

Response: yes, table 3 computed on testing performance. Figure 15 is sample visualization about two countries.

What was the source of the discrepancy in the original submission?

Response: Dear reviewer, in original submission we kept training on random partitions 80 % of countries and tried to predict for 20% on random countries. This is exact mistake were forecasting cannot be done on training and testing on randomly, some data leakage occurs.. This was mentioned clearly by previous reviewer 1 and yourselves. So again we are implemented the work on exact forecasting as said by reviewers and editors.

Now what we present is exactly have tested inference from 2024 test data.

2) The Limitations section (lines 626–632) now acknowledges that the naïve persistence baseline achieves R² = 0.887 due to strong temporal autocorrelation in annual exchange rates. While this is an improvement over the original submission, the justification for the hybrid model's practical value remains somewhat limited.

A discussion of specific forecasting scenarios where the hybrid model's incremental accuracy gains translate to meaningful economic or decision-making value

Response: Discussion is added based on given suggestion 4.1 Discussion

It would appear as though the yearly exchange rate series exhibits considerable temporal dependence; however, there will still be value in terms of both economic and decision making from the incremental accuracy gains resulting from the proposed hybrid BiLSTM + Attention + Random Forest Regressor (RFR) model. In general, even minor decreases in forecast error can produce significant financial consequences in international financial systems where exchange rates significantly influence trade flows, investment decisions and risk management practices.

An important example where improved accuracy in forecasting provides value is currency hedging in international trade. Companies which have cross border transactions frequently use either forward contracts, options or currency swaps to hedge their foreign exchange exposure. When forecast errors are reduced, companies will be able to more accurately forecast the movement of exchange rates and choose more appropriate hedging strategies. Additionally, even relatively minor improvements in predictive accuracy may reduce hedging costs or avoid unforeseen financial loss when large transaction volumes are at stake.

Additionally, another practical application of the improved forecasting accuracy provided by the hybrid model is international trade pricing and contract planning. Frequently, exporters and importers enter into long term contracts denominated in foreign currencies. More accurate forecasts of trend in exchange rates will enable companies to more accurately estimate future revenues and costs while negotiating trade agreements. Therefore, improved forecasting will enable companies to better manage their exposure to exchange rate risks and make better-informed pricing decisions in global markets.

The hybrid model will also potentially provide value for macro-economic policy monitoring. Central Banks and Financial Regulators frequently monitor the direction of movement of exchange rates when assessing the external stability of an economy. Improved forecasts may assist policy makers in identifying potential pressures on currencies, adjusting their monetary policy strategies, or evaluating the possible effects of changes in international economic conditions.

In addition to the above examples of how hybrid models may provide value, hybrid forecasting models may also have value during times of structural breaks or high levels of market volatility. Naive persistence models typically assume that the most recent observation is the best predictor of future observations. This assumption may perform reasonably well during stable economic times; however, this assumption may fail when the exchange rate dynamics suddenly change due to either financial crises, geopolitical events or major macroeconomic policy regime changes. Examples include currency devaluation, financial crisis, or a sudden change in the monetary policy regime.

The proposed hybrid architecture has been developed to capture both the temporal dependencies and non-linear relationships within the exchange rate data. The Bi-LSTM component captures the sequential patterns over time; the attention mechanism identifies the most informative time-steps; and the random forest regression captures the non-linear relationship amongst the feature extractions. Therefore, the hybrid model is better suited to identify changes in the dynamics of the exchange rates than simple persistence based models.

While the yearly data set utilized in this study smooths out many of the short-term fluctuations in the data, the results indicate that hybrid deep learning models can still utilize additional predictive information beyond simply utilizing persistence. Future studies may build upon this analysis by employing higher frequency data sets (i.e., monthly or daily exchange rate data) to further examine model performance under extreme volatility or structural economic changes.

Alternatively, an analysis of performance during periods of structural breaks or high volatility, where the naïve baseline would be expected to fail

Response: We agree it would be useful to show empirically the potential of using incremental forecasts to produce real-world benefits, especially given the high persistence in this type of data (annual exchange rates). In our updated version of the paper, we added content to both the Limitations Section and the Discussion Section to clarify how a reduction in forecasting errors could be translated into measurable economic benefits. We have updated Table 4. Table 4. Statistical Significance Test (Paired t-test on Model Residuals).

Model Comparison	Mean Residual Difference	t-Statistic	p-Value	Significance
Random Forest vs BiLSTM + Attention + RFR	-0.043	2.81	0.003	p < 0.05 (Significant)
Naïve Persistence vs BiLSTM + Attention + RFR	-0.121	3.46	0.001	p < 0.05 (Significant)
Naïve Mean vs BiLSTM + Attention + RFR	-0.390	4.12	<0.001	p < 0.05 (Significant)

The hybrid model's mean residual difference is 0.121 lower than naïve persistence baseline with a t-statistic of 3.46 and p value of .001 indicating the hybrid model is significantly better than the naïve persistence model. It should be noted that naive models perform strongly in annual exchange rate forecasting due to high temporal autocorrelation present in exchange rates. The hybrid model performs even better when compared to the naive average model. The hybrid model's mean residual difference is .390 lower than the naive average model with a t-statistic of 4.12 and a p value of less than .001 indicating the hybrid model is much more accurate than historical averages based forecasting. Therefore, the hybrid model captures important nonlinear patterns and temporal trends that are not possible using simple average-based forecasting approaches. Also, the hybrid model also shows statistically significant improvement over the Random Forest model with a mean residual difference of .043 lower than the Random Forest model, a t-statistic of 2.81 and a p value of .003 indicating the combination of bi-lstm sequential learning, attention-based feature weighting and ensemble regression provides additional predictive power beyond that of a stand-alone machine learning model.

Please add the final hyperparameter values selected by Grid Search to Table 2 or the main text to ensure full reproducibility.

RESPONSE: The hyper parameter table is added in the table 2 section 3.4 line number 311 onwards . Table 2 defines best hyperparameter of random forest regressor on GridVsearch.

Table 2: Best hyperparameter of random forest regressor on GridVsearch

	Hyperparameter	Value
0	max_depth	10
1	min_samples_leaf	1
2	min_samples_split	2
3	n_estimators	100

4) The revised manuscript still contains numerous grammatical errors and unclear phrasing. Examples include:

Lines 463–464: Inconsistent use of "Bi-LSTM" vs. "BiLSTM"

Response: All terms are used as Bi-LSTM Based on the given suggestion

Lines 623–624: The sentence beginning "The existing models all for certain country specific..." is grammatically unclear

Response: Now it's rewritten as “The existing models do not show real-time dynamics on specific countries' data, where the proposed model is a global foreign exchange forecasting model, where data from all countries are utilised to train the model. Also, proofreading is done using a standard tool.

A thorough proofreading pass is strongly recommended before final publication.

Round 3

Reviewer 3 Report

Comments and Suggestions for Authors

The authors have thoroughly addressed all major concerns raised in previous review rounds.

Article Menu

Hybrid Deep Learning Techniques Integrated with Machine Learning for Foreign Exchange Rate Forecasting

Further Information

Guidelines

MDPI Initiatives

Follow MDPI