Research on a Time Series Data Prediction Model Based on Causal Feature Weight Adjustment

Round 1

Reviewer 1 Report

In this research, the features found by the LiNGAM algorithm are used to predict stock prices after weight adjustment. The manuscript has the following problems that do not deserve to be published in the journal.

1- The research method and information of the data used in this research is not clear

2. The evaluation methods used are simple and of the same type. It is necessary to test the discrimination or probabilistic evaluation methods based on the innovative method.

3- The abstract of the research is poorly written and needs more explanations of the research results and methods

5- The interpretation of the results is weak.

It needs some editing

Author Response

Please see the attachment

Author Response File: Author Response.docx

Reviewer 2 Report

A.    Stock price predictability is a hot topic that has been addressed by many authors in the past. Work that seems to be overlooked by the authors. Just to give you a couple of examples:

1.     Goyal and Welch (2008): A comprehensive look at the empirical performance of the equity premium prediction. Review of Financial Studies. 

2.     Goyal, Welch and Zafirov (2022) have a recent update of the original paper. It is available as a Research Paper Series N° 21-85 from the Swiss Finance Institute.  

But the list of missing relevant papers is really long.

B.     Key benchmark to compare your models with: the random walk benchmark.  If you are predicting the price of the stock itself, the random walk simply states that P(t)=c+P(t-1) + error. The typical one-step ahead forecast for P(t) is just c+P(t-1). Sometimes c is set to zero. If you are forecasting one-step ahead returns, you should compare your methods with the zero forecast or with just the historical mean benchmark. The random walk benchmark will tell you if your time series are predictable or not. If they are not predictable, it probably does not make sense to compare the accuracy of methods A, B or C.

C.     Please clarify:  full sample period and evaluation period.

D.    You mention some sort of standardization so your variables will have zero mean and unit variance. If you do this with the whole sample period, you totally break the out-of-sample spirit of the predictive exercise.

E.      You never mention if your series are stationary or not.  Stock prices are traditionally modeled as having unit roots.  When working with unit roots you may encounter spurious correlations and spurious statistical relationships. That is why in time series analysis we emphasize to work with stationary transformations. I cannot see in your paper any indication of  whether the inputs to your models are stationary or not. This is a key aspect to consider.

F.      Please include charts of your series. At least of your target variables.

G.    Would your results be robust to other more common stock prices or indices? What about Dow Jones or something similar?

H.    In the forecasting literature we are very clear about the horizon we are forecasting. We report MSPE for different horizons separately. You expect higher accuracy at short horizons, and worse at long horizons. It is not clear to me if in your computations of accuracy you are including multiple horizons or not.

I.        When comparing MAPEs or MSPEs you do not report statistical significance. Maybe the seemingly more accurate methods are not statistically different from the less accurate ones.  Diebold and Mariano (1995) is a basic paper in this regard. Published in JBES.

J.       Causality and predictability may go in separate avenues. Including causal connections will not necessarily boost predictability. For references about this see Pincheira and Hardy (2019) Forecasting commodity prices with the Chilean exchange rate. (Resources Policy)

K.     Please clarify how your algorithm detects causal relationships. Please include a simple example. 

L.      It Seems to me that equation 1 may have a typo.

.  Please explain with a simple example your propose methology.

   

English can and must improve. The paper can be read, but some paragraphs are difficult to understand. 

Author Response

Please see the attachment

Author Response File: Author Response.docx

Reviewer 3 Report

We have a research paper, not a thesis, where we use the terms 'Section' and 'Chapter.' Please make adjustments accordingly.

 

Overall, the paper is well-written, and the major concepts are clearly illustrated. However, there is scope for further improvement.

 

Two main aspects that need attention are causality and the Sparcification of the relationship. Has the author explored the Granger form of causality?

 

The paper writeup should address the following issues:

 

Clarify whether the problem at hand is a forecasting problem or a regression problem.

 

If it is a forecasting problem, specify the number of steps ahead the forecast is being made.

In the related work section, examine some multivariate automated ML based forecasting algorithms (e.g., AutoAI-TS SIGMOD 2021) and compare their performance with the proposed approach.

It would be beneficial to discuss the use of Transformers for time series instead of LSTM, as Transformers are considered better in this context.

Update the related work section to include more recent research, as some of the referenced work is outdated."

None

Author Response

Please see the attachment

Author Response File: Author Response.docx

Reviewer 4 Report

This paper proposes a LSTM-base model for time series forecasting, with use of techniques of causality discovery algorithm and feature weight adjustment. I have the following comments:

1. The authors have conducted well review on RNN and CNN, but in recent years, Transformers have been considered as a strong candidate for time series, it is recommended to review such methods, e.g.

[1] Y. Lin, I. Koprinska and M. Rana, "SSDNet: State Space Decomposition Neural Network for Time Series Forecasting," 2021 IEEE International Conference on Data Mining (ICDM), Auckland, New Zealand, 2021, pp. 370-378, doi: 10.1109/ICDM51629.2021.00048.

[2] Wen, Q., Zhou, T., Zhang, C., Chen, W., Ma, Z., Yan, J., & Sun, L. (2022). Transformers in time series: A survey. arXiv preprint arXiv:2202.07125.

2. How the 33 initial features were selected, based on automatic methods or were selected manually?

3. The section 3.3 can be moved to section 2 since background of LSTM should in related work.

4. The proposed methods perform well on Chinese stock datasets, I am interested if this method is also appliable on other national data, e.g. the exchange rate data in this paper:

[1] Y. Lin, I. Koprinska and M. Rana, "SSDNet: State Space Decomposition Neural Network for Time Series Forecasting," 2021 IEEE International Conference on Data Mining (ICDM), Auckland, New Zealand, 2021, pp. 370-378, doi: 10.1109/ICDM51629.2021.00048.

Author Response

Thank you for your questions and our answers are in the attachment.

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

The manuscript was partly revised based on the given comments, only the results need to be expressed in more figures to give the reader a better understanding of the research.

it needs some editing

Author Response

Thanks to your suggestion, we have made the days of data clearer in Section 4.

Reviewer 2 Report

Comments for author File: Comments.pdf

Can be improved. 

Author Response

Thank you for your suggestions and help in revising this article!

Author Response File: Author Response.pdf

Reviewer 3 Report

Overall paper looks in good shape.

Author Response

Thank you for your suggestions and help in revising this article!