MySTOCKS: Multi-Modal Yield eSTimation System of in-prOmotion Commercial Key-ProductS

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The article explores the issues of forecasting the availability of agricultural goods to meet consumer demand. The topic of the work is important and relevant for the chosen field. The authors conducted a study on the application of the MySTOCKS system for effective inventory management in the agri-food and large-scale retail environment. The MySTOCKS system uses an advanced deep learning framework to improve operational response to dynamic requirements of the agri-food supply chain. Overall, the article is interesting and has important results. All figures and tables in the article have appropriate references. 

Notes for improving the article:

Modern literature is used in the Introduction Section and the Related Works Section, but only 3 literary sources for the years 2016-2021 are used in the Introduction Section, and literary sources for the years 2019-2021 are used in the Related Works Section. Newer literature for the years 2022-2024 should be added to these Sections. When citing, you should use the notation [2,3] in line 50, [8-11] in line 142, [4,5] in line 154, [6,7] in line 160, [10,11] in line 183.

At the end of the Introduction Section, you should give a brief description of what will be considered in each of the following Sections.

If in Figure 2 and Figure 3 the authors showed part of the image from Figure 1, then for better detailing, Figure 2 and Figure 3 should be enlarged to the allowable text width.

It is necessary to check that the formulas contain explanations for all used parameters, for example, c_i in formula (1) from line 325, l_x in formula (2) from line 328,. n, t in formula (3) from line 332, Q, K, V in formula (4) from line 337.

Lines 463-468: “Neuroscience research has identified two types…” It is probably worth giving references to literary sources.

In Section 5 Conclusions, the information is repeated. So, why give the indicators of the MySTOCKS system in lines 618-619, if then they are again compared with other methods in lines 642-655.

 

Author Response

1. Modern literature is used in the Introduction Section and the Related Works Section, but only 3 literary sources for the years 2016-2021 are used in the Introduction Section, and literary sources for the years 2019-2021 are used in the Related Works Section. Newer literature for the years 2022-2024 should be added to these Sections. When citing, you should use the notation [2,3] in line 50, [8-11] in line 142, [4,5] in line 154, [6,7] in line 160, [10,11] in line 183.

 

Authors: Thanks for the suggestions. Modern literatures were added, and citation were formatted in correct manner.

 

2. At the end of the Introduction Section, you should give a brief description of what will be considered in each of the following Sections.

Authors: Thanks for the suggestion. The Introduction section was improved with a brief description of the following sections.

3. If in Figure 2 and Figure 3 the authors showed part of the image from Figure 1, then for better detailing, Figure 2 and Figure 3 should be enlarged to the allowable text width.

 

Authors: Thanks for the suggestions. The figures were improved.

 

4. It is necessary to check that the formulas contain explanations for all used parameters, for example, c_iin formula (1) from line 325, l_x in formula (2) from line 328,. n, t in formula (3) from line 332, Q, K, V in formula (4) from line 337.

Authors: Thanks for the suggestion, formulas were improved

5. Lines 463-468: “Neuroscience research has identified two types…” It is probably worth giving references to literary sources.

 

Authors: Thanks for the suggestion, reference were added.

6. In Section 5 Conclusions, the information is repeated. So, why give the indicators of the MySTOCKS system in lines 618-619, if then they are again compared with other methods in lines 642-655.

 

Authors: Thanks for the suggestion. Conclusion section were improved.

Reviewer 2 Report

Comments and Suggestions for Authors

The paper introduces the MySTOCKS system, a dual-transformer-based architecture designed for inventory forecasting, specifically addressing out-of-stock (OOS) and surplus-of-stock (SOS) challenges. The proposed solution is innovative, leveraging transformer models, Elastic Weight Consolidation (EWC), and multi-modal data processing for both standard and promotional scenarios. The manuscript is well-structured, covering the problem context, methodology, results, and comparisons comprehensively. However, some areas require clarification and further elaboration.

Major Concerns

1.       The authors state that data cannot be shared due to company policy and patent submission. This restricts the reproducibility of results and limits validation by independent researchers. Consider providing synthetic or anonymized datasets for transparency.

2.       While the paper mentions high accuracy (up to 93.8%), details on test conditions, cross-validation protocols, and dataset splitting are limited. Including these would strengthen the validity of the reported results.

3.       TR2’s forecasting horizon of 45 days is impressive but lacks justification regarding its practical feasibility in dynamic retail settings. Additional case studies or real-world deployment examples would be beneficial.

4. Provide additional real-world deployment examples or validation results from industrial partners to strengthen the practical applicability of the system.
5. Clarify how the system ensures GDPR compliance, particularly with sensitive promotional data.

6.       The paper compares the MySTOCKS system against traditional methods (e.g., LSTM, TCN). Including results from recent transformer-based methods would provide a more thorough benchmark.

7.       Certain technical terms (e.g., "Markovian dynamics," "softmax layers") are introduced without adequate context for a broader audience. Brief explanations or a glossary would enhance accessibility.

8.       Figures 2 and 3 illustrating the TR1 and TR2 architectures lack sufficient detail in captions. Consider elaborating on their components and relevance.

9.       Acronyms like OOS, SOS, EWC, and GDPR are used extensively. While most are defined early, ensuring consistent usage throughout would improve readability.

10.  The manuscript contains minor formatting inconsistencies, such as incomplete references . Address these before publication.

Author Response

1. The authors state that data cannot be shared due to company policy and patent submission. This restricts the reproducibility of results and limits validation by independent researchers. Consider providing synthetic or anonymized datasets for transparency.

Authors: We appreciate the reviewer’s concern regarding reproducibility and validation. However, due to strict company policies and the constraints imposed by the patent submission process, we are unable to share any portion of the dataset, including synthetic or anonymized versions. To ensure transparency, we have provided a description of the dataset structure (rows 527-543), feature selection process, and data preprocessing pipeline in the manuscript (Section 3). Moreover, the methodologies employed in this study are well-documented and can be applied to similar publicly available datasets, allowing independent validation of the proposed approach. We hope this explanation clarifies our position and appreciate the reviewer’s understanding of these limitations.

2. While the paper mentions high accuracy (up to 93.8%), details on test conditions, cross-validation protocols, and dataset splitting are limited. Including these would strengthen the validity of the reported results.

Authors: We appreciate the reviewer’s suggestion regarding the validation of our results. Below, we clarify the evaluation procedure used in our study.

• Dataset Splitting: The dataset was split into 70% training, 15% validation, and 15% testing. This ensures that model tuning was conducted on a separate validation set before final evaluation on an unseen test set, reducing the risk of overfitting.
• Cross-Validation: We employed a 3-fold cross-validation approach to enhance the robustness of our results. This method ensures that the model’s performance is not dependent on a single train-test split but is instead averaged over multiple iterations, improving generalization.
• Test Conditions: The model was evaluated under two distinct conditions:

• Standard inventory forecasting (TR1): Tested on regular sales patterns without promotional influence.
• Promotional forecasting (TR2): Evaluated over extended forecast windows (up to 45 days), ensuring its ability to predict stock levels under high-demand periods.
  • Performance Metrics: The reported accuracy of up to 93.8% was computed using standard classification metrics (accuracy, sensitivity, and specificity), comparing predicted vs. actual stock levels across multiple runs to ensure consistency.

We recognize the importance of detailing these validation steps and will incorporate the missing information in the manuscript. Thank you for your valuable feedback.

3. TR2’s forecasting horizon of 45 days is impressive but lacks justification regarding its practical feasibility in dynamic retail settings. Additional case studies or real-world deployment examples would be beneficial.

Authors: We appreciate the reviewer’s comment regarding the forecasting horizon of 45 days for TR2. The selection of this time frame was motivated by practical considerations observed in real-world retail operations.

• Alignment with Retail Planning Cycles: In large-scale retail, promotional campaigns are typically planned several weeks in advance, requiring inventory forecasting that extends beyond short-term horizons. The 45-day period was chosen to align with standard promotional planning cycles, which often include pre-promotional stock buildup, active promotional sales, and post-promotion stock clearance.
• Capturing Promotional Dynamics: Unlike standard demand forecasting, promotional inventory planning requires anticipating both peak sales during promotions and potential surplus issues at the end of the period. A shorter horizon may not provide sufficient foresight to optimize stock levels across the entire promotional lifecycle.
• Empirical Model Performance: Despite the extended forecasting window, our experimental results show that TR2 maintains high accuracy (89%), sensitivity (88%), and specificity (90%), demonstrating its reliability in predicting stock dynamics over this time frame.

We recognize the value of supplementing our study with real-world case studies, and we are actively exploring future opportunities to validate our approach in operational retail environments.

 

4. Provide additional real-world deployment examples or validation results from industrial partners to strengthen the practical applicability of the system.

 

Authors: We appreciate the reviewer’s suggestion regarding real-world deployment examples. The MySTOCKS system has been validated using realistic retail inventory datasets, reflecting industry challenges in stock forecasting during both standard and promotional periods. While specific deployment case studies from industrial partners cannot be disclosed due to confidentiality agreements, the methodology and results presented in the manuscript demonstrate the practical applicability of the system. Additionally, we are exploring future opportunities for large-scale deployments, which could be included in subsequent studies. Thank you for your valuable feedback.

 

 

5. Clarify how the system ensures GDPR compliance, particularly with sensitive promotional data.

 

Authors: We ensure GDPR compliance through the following measures:

• No Personal Data: The system processes only aggregated product and sales data, excluding personally identifiable information (PII).
• Data Minimization: Only essential inventory and promotional data are collected and stored.
• Security Measures: Encryption, access controls, and audit logs ensure data protection.
• Regulatory Compliance: The system aligns with GDPR Articles 5, 25, and 32, ensuring lawful, secure, and privacy-focused data processing.

Thank you for your valuable suggestion. The necessary clarifications have been added in lines 521-524

6. The paper compares the MySTOCKS system against traditional methods (e.g., LSTM, TCN). Including results from recent transformer-based methods would provide a more thorough benchmark.

Authors:  We appreciate the reviewer’s suggestion regarding additional benchmarks with recent Transformer-based methods. Our comparison focused on widely used traditional architectures (LSTM, TCN, and FCN) to demonstrate the improvements introduced by MySTOCKS in inventory forecasting. While we acknowledge the relevance of newer Transformer-based models such as Temporal Fusion Transformer (TFT) or Informer, our system differs significantly as it integrates dual forecasting modules (TR1/TR2), domain adaptation, and Elastic Weight Consolidation (EWC) to enhance learning stability and mitigate catastrophic forgetting. Given these structural differences, a direct comparison with other Transformer-based models is not straightforward.

7. Certain technical terms (e.g., "Markovian dynamics," "softmax layers") are introduced without adequate context for a broader audience. Brief explanations or a glossary would enhance accessibility.

Authors: We appreciate the reviewer’s suggestion regarding the clarity of technical terms. To improve accessibility, we have added brief explanations for key concepts such as "Markovian dynamics" (rows 372-373) and "softmax layers" (row 282) where they are first introduced in the manuscript. We believe these clarifications will enhance readability while maintaining the technical rigor of the paper. Thank you for your valuable feedback

8. Figures 2 and 3 illustrating the TR1 and TR2 architectures lack sufficient detail in captions. Consider elaborating on their components and relevance.

Authors: Thanks for the suggestion. The captions were improved.

9. Acronyms like OOS, SOS, EWC, and GDPR are used extensively. While most are defined early, ensuring consistent usage throughout would improve readability.

Authors:  We appreciate the reviewer’s feedback regarding acronym usage. We have reviewed the manuscript to ensure that all acronyms (e.g., OOS, SOS, EWC, and GDPR) are clearly defined at their first occurrence and consistently used throughout the text. This refinement enhances readability while maintaining technical precision. Thank you for your valuable suggestion.

 

10. The manuscript contains minor formatting inconsistencies, such as incomplete references. Address these before publication.

Authors:  We appreciate the reviewer’s feedback regarding formatting inconsistencies and incomplete references. We have carefully reviewed the manuscript and corrected these issues to ensure consistency and completeness. Thank you for your helpful suggestion.

 

 

 

We thank the reviewer for the suggested changes, thanks to his comments we had the opportunity to improve our article.

Round 2

Reviewer 2 Report

Comments and Suggestions for Authors

The authors have addressed my concerns well.