Generalized Loss-Based CNN-BiLSTM for Stock Market Prediction

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The article introduces a new method called generalized loss CNN-BiLSTM (GL-CNN-BiLSTM), which dynamically calculates the cost of errors based on the difficulty of the data, aiming to minimize costly errors that lead to trading losses. Experimental results demonstrate that while GL-CNN-BiLSTM shows no significant improvement in accuracy compared to other models, it achieves the highest rate of return on test data from the Shanghai, Hong Kong, and NASDAQ stock exchanges. The paper is interesting and tackles a timely topic. However, I have several comments to improve the work and bring it closer to publishability.

1. The discussion of the contribution in the introduction could be written in plain text rather than as enumerated points. Furthermore, it would be valuable to link the paper to the most relevant works in the field to better illustrate how the authors address the research gap. It would be beneficial to add one more paragraph at the end of the introduction outlining the structure of the remainder of the article.

2. Would the readers know what “CNN-BiLSTM” in the title means? If not, perhaps it is better to reformulate it to avoid confusing abbreviations.

3. I cannot see much discussion on the validation and tuning of the models. How is the sample split into training, validation, and testing? Overall, the employed models require several different parameter assumptions. The authors should disclose them properly to ensure the reproducibility of the research.

4. The data description in Section 5 is very scarce. The authors should dedicate more space to sample descriptions, descriptive statistics, preparation procedures, and cleaning. The authors write, “we removed the missing value, and then z-score was implemented to detect the outliers and standardization.” Is the z-score standardization done monthly? Why do the authors choose the z-score over, for example, rank-based mapping as in Cakici et al. (2023)?

5. The discussion of the results is absolutely insufficient. The authors should dedicate much more space and effort to present their findings in Section 7. Most of the results are left undiscussed.

6. Several of the observations are quite interesting, and the authors should consider exploring them further. For example, while the differences in Sharpe ratios are substantial, the accuracy measures are almost identical. What is the source of this discrepancy? Furthermore, why do some models perform much better in terms of their Sharpe ratios than others?

7. I would be interested in seeing at least some discussion of practical considerations, such as trading costs.

8. To improve the positioning of the paper, the authors may consider linking it more strongly in Section 2 with works on applying deep learning methods to predictions of aggregate stock market returns, such as Chong et al. (2017) and Zhou et al. (2023).

9. The authors should make sure that their exhibits are self-contained. Currently, most of the symbols, etc., are left unexplained.

References:

Chong, E., Han, C., & Park, F. C. (2017). Deep learning networks for stock market analysis and prediction: Methodology, data representations, and case studies. Expert Systems with Applications, 83, 187-205.

Cakici, N., Fieberg, C., Metko, D., & Zaremba, A. (2023). Machine learning goes global: Cross-sectional return predictability in international stock markets. Journal of Economic Dynamics and Control, 155, 104725.

Zhou, X., Zhou, H., & Long, H. (2023). Forecasting the equity premium: Do deep neural network models work? Modern Finance, 1(1), 1–11. https://doi.org/10.61351/mf.v1i1.2

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

ijfs-3053980

Generalized Loss-Based CNN-BiLSTM Stock Market Prediction

The authors method a generalized loss CNN-BiLSTM (GL-CNN-BiLSTM), where the cost of each data can be dynamically calculated based on the characteristics of the data. The authors verify the effectiveness of their method on Shanghai, Hong Kong, and NASDAQ stock market data. Overall, while the combination of CNNs and LSTMs can yield powerful models capable of capturing both spatial and temporal features, the associated weaknesses highlight the need for careful consideration, and I would suggest the authors discuss the following:

1). In terms of computational complexity, the combination of CNN and LSTM layers can result in a model that is computationally expensive and memory intensive, requiring significant hardware resources for training and inference.

2). Give 1), training such combined models can be time-consuming due to their complexity and the large amount of data required. Please explain and discuss this issue.

3). Is the combined model introducing a large number of hyperparameters (e.g., number of LSTM units and so on) making the tuning process more complicated and time-consuming. Please specify. If yes, then, there is the problem of overfitting.

4) Also, the LSTM component captures long-term dependencies, but the exact nature of these dependencies is encoded in the model weights and hidden states, which are not straightforward to interpret. In addition, the final prediction is the result of a complex interplay between the features extracted by the CNN and the temporal dependencies captured by the LSTM. How can the analyst decompose this combined decision-making process to explain why the model is confident about forecasting?

5). When the model makes incorrect predictions, diagnosing and fixing the issues is harder without clear interpretability. Please provide a discussion/guideline as to which part of the model (CNN or LSTM) contributed to the error.

6). Please update the literature.

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

The authors have resolved all my suggestions. I have no further comments but one: I find the idea of including a table with abbreviations in the introduction a bit bizzare. Perhaps the authors may just briefly explain them in a footnote or move this table to an appendix?

Reviewer 2 Report

Comments and Suggestions for Authors

Thank you for the revised version