Final peer review

[pdarbha]

This project is attempting to predict the price of Bitcoin using a variety of features that are indicative of the economy such as the S&P, interest rates, inflation rate, etc. as well as the price of GPUs since that is very important for the mining of Bitcoin.

Things I liked:

1. I liked that they thoroughly explain the features they used, why they are important, and where they got the data from. As someone who was not familiar with the process of pricing Bitcoin, I found the explanations very helpful in understanding the rest of the paper.
2. I liked the variety of visualizations they used which showed different aspects of the data.
3. I liked that they explained the models and the functions. Even though we have taken the class and are aware of what the different terms, such as RBF and polynomial kernel, mean, it is good to have these explanations for general readers.

Things I did not like:

1. While the visualizations were helpful, there may have been a bit too much. It was getting a little hard to read the report sometimes because it kept being interrupted by the visualizations. Using fewer and maybe making them bigger would have helped.
2. I would have appreciated a little more on why the different models performed better or worse than each other based on the kind of data.
3. I would have liked to have seen different kinds of predictions like leaving out internal points and trying to predicting those. Maybe the final 20% was much different than the distribution as a whole.

[TLI-1994]

Thank you for the suggestions. Here are some comments:

1. While the visualizations were helpful, there may have been a bit too much. It was getting a little hard to read the report sometimes because it kept being interrupted by the visualizations. Using fewer and maybe making them bigger would have helped.

The visuals are added according to the contents and all of them make sense to the text. It is hard to get rid of any of them. Please indicate which figure is believed to be redundant so that we can double-check it. We also used hyperlinks pointing to figures, and we hoped the hyperlinks helped with the consistent reading experience.

2. I would have appreciated a little more on why the different models performed better or worse than each other based on the kind of data.

This is definitely a very good suggestion. We have indeed included some brief discussions on this, but admittedly we need more to make clear sense. A thorough review of the ML algorithms is required to fulfill this requirement given that we are dealing with multiple highly integrated algorithms. We find it beyond the scope of this report since we have to present many other aspects of our work as well. But still, this is a valuable suggestion.

3. I would have liked to have seen different kinds of predictions like leaving out internal points and trying to predicting those. Maybe the final 20% was much different than the distribution as a whole.

Thank you for this suggestion. We could've included our opinions on this specific topic but the length of the report is limited. Here is a brief version of this discussion: first of all, these ideas have been investigated, and obviously there is more than one issue with this idea: 1. For such a time-series problem, it makes no sense predicting (or more precisely, fitting) internal points. If we have to "predict" the internal points, the error, as we can confidently say, will be very small. But how does this kind of model help people to invest anyway? Admittedly, doing such analysis can provide more insights about which factor is the most influential for bitcoin pricing in history, but the report is, unfortunately, focused on predicting. 2. the authors are aware the final 20% data has a different distribution, that is why we did stationarity analysis and did the re-formalization. We hoped the readers would figure out the purpose of that section following the linear models, but it seems we have to underscore it a little bit more for the readers to notice.