Predicting Price Changes in Ethereum (2017) [pdf]
cs229.stanford.edu
Predicting Price Changes in Ethereum (2017) [pdf]
1–10 of 42 posts
Re: Predicting Price Changes in Ethereum (2017) [pdf]
#2You should predict returns or log returns - not the actual price series - because of non-stationarity.
For a good reference, see Hayashi.
Re: Predicting Price Changes in Ethereum (2017) [pdf]
#3Interesting that even the most naive methods still have >50% accuracy. Also interesting that the best method was better able to predict downward moves than upward moves, during a bull market. Any intuitive reason for this?
Is there some reason the study doesn't include the post-December 2017 bear market?
Re: Predicting Price Changes in Ethereum (2017) [pdf]
#4I can vouch for the DNN results. I took data from Kaggle and tried training networks with different approaches. Was not able to get prediction any better than 54% (which I am pretty sure would be biased as well). Later on while inspecting data and looking at dips and rises turns out external events (govts banning crypto or a price crash due to other reasons) was causing price drops or surge, my conclusion was if you need to predict anything you need to be aware of these external factors.
Re: Predicting Price Changes in Ethereum (2017) [pdf]
#5> The primary dataset consists of the price of Ether sampled at approximately one-hour intervals between August 30, 2015 and December 2, 2017.
As far as I am aware, automated trading often works with much shorter intervals, in the milliseconds range. The traditional stock trading industry also does not only look at the price history, but also at planned buy and sell orders in the books and even news and social media sentiment analysis. Perhaps prediction methods that also use these strategies would produce better results, especially given that the crypto market has been in such a chaotic flux between 2015 and 2017.
Re: Predicting Price Changes in Ethereum (2017) [pdf]
#6> The primary dataset consists of the price of Ether sampled at approximately one-hour intervals between August 30, 2015 and December 2, 2017. As far as I am aware, automated trading often works with much shorter intervals, in the milliseconds range. The traditional stock trading industry also does not only look at the price history, but also at planned buy and sell orders in the books and even news and social media…
There doesn't seem to be enough frequency in crypto trading for meaningful info at millisecond resolution.
Re: Predicting Price Changes in Ethereum (2017) [pdf]
#7> The primary dataset consists of the price of Ether sampled at approximately one-hour intervals between August 30, 2015 and December 2, 2017. As far as I am aware, automated trading often works with much shorter intervals, in the milliseconds range. The traditional stock trading industry also does not only look at the price history, but also at planned buy and sell orders in the books and even news and social media…
There doesn't seem to be enough frequency in crypto trading for meaningful info at millisecond resolution.
minutes or seconds would be more useful than one hour intervals.
Re: Predicting Price Changes in Ethereum (2017) [pdf]
#8This is OK as a final project in an intro econometrics class, but nothing else.
Re: Predicting Price Changes in Ethereum (2017) [pdf]
#9Earlier quoted context omitted.
There doesn't seem to be enough frequency in crypto trading for meaningful info at millisecond resolution.
minutes or seconds would be more useful than one hour intervals.
This is assuming no trading fees and highly liquid exchanges with low spreads, otherwise the profit from these micro-movements could be eaten up. I see no problem with automatic trading over periods of hours if you're getting >60% accuracy.
Re: Predicting Price Changes in Ethereum (2017) [pdf]
#10Interesting that even the most naive methods still have >50% accuracy. Also interesting that the best method was better able to predict downward moves than upward moves, during a bull market. Any intuitive reason for this? Is there some reason the study doesn't include the post-December 2017 bear market?
> Is there some reason the study doesn't include the post-December 2017 bear market?
Now, that would be a decent out of sample test, wouldn't it?
Also, keep in mind that at hourly resolution, with a highly autocorrelated dataset, 60% is not much better than coin toss.
I would like to see if any of the methods they used would outperform just using what happened in the previous period as one's prediction.