Ultra-Short-Term Continuous Time Series Prediction of Blockchain-Based Cryptocurrency Using LSTM in the Big Data Era
Round 1
Reviewer 1 Report
Dear colleague(s),
The study Ultra-short-term continuous time series prediction of block- 2 chain-based cryptocurrency using LSTM in the Big Data Era sounds interesting and it is mostly well-organizes.
There 3 queries that should be answered and included which will influence higher quality of this study:
1. Why you didnt concern market efficiency ? This is crucial point in time series forecasting for financial data. Even more you should discuss and include Random Walk as your topic deal with this part of market efficiency.
2. How you are sure that forecasting method is appropriate? I do not see any results of forecasting errors? You should calculate MAPE, MAE, RMSE or some other errors to get predicting power of the model.
3. Lastly, there is no discussion of obtained results and link with other literature with similar methodology.
Author Response
Please see the attachment.
Author Response File: 
Author Response.docx
Reviewer 2 Report
The results on BTC or ETH price prediction using LSTM could be made stronger by extending the dataset, especially including any major event . Due to a small dataset it is difficult to establish a price prediction model.
You should also state the accuracy % of a model.
Furthermore, I did not find reference [24] , and [25] in the content of a paper
Overall, the research on BTC and ETH price using a minute level dataset is good but needs improvement in results.
Author Response
Please see the attachment.
Author Response File: Author Response.docx
Reviewer 3 Report
Following is some suggestions
1. Authors could provide summary statistics of their dataset.
2. Instead of a generic diagram of LSTM, the architecture of created model used could be included.
3. Few references are not cited within main text of manuscript. For example reference number 16.
4. Figure 4 explanation of flowchart of system must avoid using the rounded numerals like (1) in text lines 199- 220.
5. Before conclusions, the performance metrics of model can be reported and discussed in section 5. At present it mentions the real values only in tables. Standard values like RMSEA, F1-score, etc. can be included.
Author Response
Please see the attachment.
Author Response File: Author Response.docx
Reviewer 4 Report
Before this manuscript can be considered for publication the Authors need to take the following remarks into consideration:
- The investment strategy described works in such a way that the LSTM network predicts whether the next minute closing price will be higher or lower than the previous minute's close and makes buying decisions based on that. If not, they are waiting, and if they have already bought, they check if the closing price in the next minute will be lower, as they sell it. The question here is how quickly a decision to buy or sell is made. The decision is made based on the previous 200 minute candles and the last price is the closing price. But even if the network makes a decision in a few ms (although in 2.2 they wrote that "However, there is a disadvantage in that the operation speed is slow. 120"), the transaction will not be concluded at the closing price of the previous candle, only during for the next minute. And it looks like they are testing their system on historical data as if they concluded a trade at the closing price of the last analyzed minute, and not the price of the next minute candle. Added to this are the costs of slippage and commission. To judge if it works, you'd have to play it live. In the assessment of the trading algorithm, in addition to the profit at the end, there is also a significant maximum loss along the way - maximal drawdown. Even if on average we make a profit at the end, but lost a lot along the way, we may not have the resources to continue playing.
- The second important point is that they treat the LSTM network like a black box. They wrote nothing about the architecture of the LSTM network they use: how many layers, how many neurons, how many parameters does it match? Does their network re-learn each time from the previous 200 mins or is it already learned on the dataset and only drop the result based on the 200 mins?
- In several places, cryptocurrencies are referred to as stock. Eg line 187 "The researcher wants to make a time series forecast for two stocks." "and go through the process of retrieving the following 200 cryptocurrencies, and repeat this process.", line 219.
- "Following this prediction, the actual trading of 80 pairs (80 purchasing, 80 selling) was carried out, and the number of trading..." I figured it was about 80 buy and sell orders. But a trading pair is different.
- Lines 73 and 89 about Satoshi Nakamoto are repeated.
- The content of lines 240-245 are vague and non-grammatical: "As shown in Figure 5, from 13:10 on November 22, 2021, to 12:50 on November 23, 240 2021, through 80 pairs of bitcoin trading, returning from an uptrend to a downtrend as follows, buy a sell value 62.5% of the price was higher than the price, and 78.75% of the next buy value was lower than the sell value. Here, it was confirmed that, based on the average value of buying and selling (50%), positive points for gain take advantage of + 12.5% ​​(62.5-50) in selling and + 28.75% (78.5-50) in buying. This proved that it affects the total gain (+ 4.44%). "
- The information "Although the global economic market has recently been slowing due to COVID-19, the blockchain and cryptocurrency market is rapidly growing as part of the fourth industrial revolution. In addition, as interest rates are reduced due to the global economic market's slow growth, making it challenging to prepare for retirement or raise funds through bank deposits, investment in various financial products, including stocks and cryptocurrencies, is rapidly increasing. " from the beginning of Introduction is no longer available.
The world economy continues to move quite fast after rebounding from the covid trough, but there are already some signs of a slowdown. The crypto and stock indices peaked in November 2021 and now we have a bear market. Interest rates are no longer zero. In the USA, the index is 3.25%
- Finally, reference to the relevant literature on recent developments on the cryptocurrency market is here very scanty. There is only Applied but not much of Sciences. For instance the work by M. Watorek et al on "Multiscale characteristics of the emerging global cryptocurrency market" in Physics Reports 901 (2021) 1–82; https://doi.org/10.1016/j.physrep.2020.10.005, documents several nonlinear temporal correlations (like multifractal ones) in the cryptocurreny market. Existence of such correlations indicates departures from the conventional Efficient Market Hypothesis and thus some potential for successful predictions. Such a reference should be given here.
Author Response
Please see the attachment.
Author Response File: Author Response.docx
Round 2
Reviewer 1 Report
Unfortunately I do not see a significant improvement of this work. Especially since the authors did not present the results of forecasting errors, regardless of the time period they forecast. Second, although the authors discussed some papers in the literature section, there is no connection between the results of this study and the results of the discussed papers.
Author Response
Please see the attachment.
Author Response File: Author Response.docx
Reviewer 4 Report
I recommend to take my original remarks more seriously into consideration. Just prevailing promises that it will be done in future research is not satisfactory.
Author Response
Please see the attachment
Author Response File: Author Response.docx
Round 3
Reviewer 1 Report
Dear authors,
Your paper is improved. Please make minor changes before final acceptance:
In forecasting studies, authors are presenting errors in table form (and discuss different errors). Please see benchmark study:
Are CDS spreads predictable during the Covid-19 pandemic? Forecasting based on SVM, GMDH, LSTM and Markov switching autoregression. Expert Systems with Applications, 116553
You can also highlight in limitation why you did not use any LSTM optimization!
In discussion section, you did not connect your results with the previous mentioned studies (from literature). What you received different? What new your study offers?
Author Response
Please see the attachment.
Author Response File: Author Response.docx
Reviewer 4 Report
The revisions now introduced are acceptable and thus the manuscript can be considered for publication.
Before that, however, I recommend to avoid a strange looking form as in line 164 '...the Researcher was skeptical...' lines 247 'The researcher wants 247 to make...' as it makes not clear who here is to be referred to as a 'researcher'. The author, a reader or someone elase?
Also, the singular form "I was...", line 207, looks inappropriate as there are two authors.
In ref.[15] the volume number (901) is missing.
Author Response
Please see the attachment.
Author Response File: Author Response.docx
