Research on Quantitative Investment Strategies Based on Deep Learning

Round  1

Reviewer 1 Report

1.      The format should be checked again, e.g. Fan and Palaniswami [1616] , and Tay and Cao [1212] on Page 3.

2.      The figures are not clear enough.

3.      How to determine the parameters in LSTM and LSTM-SVM? It should be presented.

4.      How to implement the in LSTM and LSTM-SVM? Which software? It should be presented.

5.      The styles in references should comply with the journal.

Author Response

List of Responses
Dear Editors and Reviewers:
Thank you for your letter and for the reviewers’ comments concerning our manuscript entitled “Research on Quantitative Investment Strategy Based on Deep Learning” (ID:431241). Those comments are all valuable and very helpful for revising and improving our paper, as well as the important guiding significance to our researches. We have studied comments carefully and have made correction which we hope meet with approval. Revised portion are marked in yellow in the paper. The main corrections in the paper and the responds to the reviewer’s comments are discussed in the attachment WORD file.

Author Response File: Author Response.pdf

Reviewer 2 Report

I find the text as a quality one. I evaluated especially topicality, scientific level, value added and approach of the authors. The work of the authors was very systematic. They prepared very clear and sufficient literature review. They described fully used dataset and methods. The results are also clear and understandable. If I can judge the topic is up to date. However, we can find a lot of papers that were focused on quantitative investment strategies we should admit that so far results are not sufficient. A few of the papers are interested in using deep learning. The authors have mentioned them mainly in literature review. Authors suggest new method of deep learning based on combination of Long Short-Term Memory network and Support Vector Machine. They use abbreviation LSTM-SVM for the suggested method. They proved that the suggested LSTM-SVM method provides higher performance than other methods (especially LSTM). I appreciate also their focus on time for computation. It is necessary because the trade of 50 ETF options (the case of the paper) is made in 15 minutes` period. Generally, the scientific level is quite high. The results that are offered by the authors are clear and valuable. 

Author Response

Response 1:

Response 1: We are very sorry for providing not sufficient results. To make more comparison between different deep learning algorithms. We add random forest to do regression in this article and compare the prediction result with the result of LSTM. We find that the result of Random Forest is worse than the result of LSTM. Therefore, we choose LSTM to combine with SVR. Then, we find the result of LSTM-SVR is the best compared with Random Forest and LSTM.

Here are the introduction and result of Random Forest:

n this article and compare the prediction result with the result of LSTM. We find that the result of Random Forest is worse than the result of LSTM. Therefore, we choose LSTM to combine with SVR. Then, we find the result of LSTM-SVR is the best compared with Random Forest and LSTM. The introduction and result of Random Forest has been discussed in the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

The paper proposes a hybrid deep learning - SVM method for forecasting the option market. The paper is quite well-written and takes into account several important details on the financial side. However, I have some important concerns on the machine-learning side of this application.

1) Traditional SVMs are classification models. It is not entirely clear how they were applied to forecasting. Apart from the input to the network, the targets should be also clearly defined, along with any pre-processing applied to them (e.g., standarization).

2) Which is the loss function used for training the LSTM. Is it the mean squared error? Also, the authors state "represents the Loss Function, which is the same as the traditional back propagation neural networks". However, to the best of my knowledge, there is no standard loss function used in the back propagation algorithm (the back-propagation algorithm is used with different loss functions). Perhaps, the authors mean the mean squared error (since this is the loss used to introduce the back-prop algorithm in many textbooks), but this should be clarified.

3) Only the prediction of the LSTM was used as feature? Why not directly use the hidden state vector, which would probably provide more information? 

4) The hyper-parameters used for the models (e.g., dimensionality of the hidden state for LSTM, learning rate, optimizers, batch size, etc.) should be reported.

5) In the deep learning community, this kind of application, i.e., just using the output of a deep learning model to train another classifier would raise several concerns. It is more common nowadays to directly train the resulting model directly with the objective of the final classifier. However, given the particular application and the audience of the journal, I think this is OK.

6) There are several recent related works on using deep learning and machine learning models to tackle these kind of tasks, which apply the models in a more structured way. The authors should refer to these works:

Chen, Kai, Yi Zhou, and Fangyan Dai. "A LSTM-based method for stock returns prediction: A case study of China stock market." Big Data (Big Data), 2015 IEEE International Conference on. IEEE, 2015.

Tsantekidis, Avraam, et al. "Using deep learning to detect price change indications in financial markets." Signal Processing Conference (EUSIPCO), 2017 25th European. IEEE, 2017.

Kercheval, Alec N., and Yuan Zhang. "Modelling high-frequency limit order book dynamics with support vector machines." Quantitative Finance 15.8 (2015): 1315-1329.

Tsantekidis, Avraam, et al. "Forecasting stock prices from the limit order book using convolutional neural networks." Business Informatics (CBI), 2017 IEEE 19th Conference on. Vol. 1. IEEE, 2017.

Passalis, Nikolaos, et al. "Temporal Bag-of-Features Learning for Predicting Mid Price Movements Using High Frequency Limit Order Book Data." IEEE Transactions on Emerging Topics in Computational Intelligence (2018).

Author Response

List of Responses
Dear Editors and Reviewers:
Thank you for your letter and for the reviewers’ comments concerning our manuscript entitled “Research on Quantitative Investment Strategy Based on Deep Learning” (ID:431241). Those comments are all valuable and very helpful for revising and improving our paper, as well as the important guiding significance to our researches. We have studied comments carefully and have made correction which we hope meet with approval. Revised portion are marked in yellow in the paper. The main corrections in the paper and the responds to the reviewer’s comments are discussed in the attachment WORD file.

Author Response File: Author Response.pdf

Round  2

Reviewer 3 Report

The authors fully addressed all my concerns and the paper can be accepted.