In this article, you will learn why one should take insights on seasonality of sentiment in Bitcoin markets with a grain of salt—even if they look interesting.
Which day is the best day to buy Bitcoin? This question pops up regularly on social media and blogs. There are many reasons why these analyses should be questioned. Still, it leads us to ask on which days of the week people express more positive or negative sentiments towards Bitcoin on Twitter, based on Augmento data.
First things first: normalization
We used a sample from daily Twitter sentiment data related to Bitcoin ranging from 2016-11 to 2019-03. We normalized the data. This is necessary since different periods are characterized by the diverging volume of tweet activity which, if not normalized, would render a comparison meaningless.
First of all, we created relative sentiment numbers. This is done by dividing the absolute sentiment numbers related to bitcoin by total mentions of bitcoin for a given period.
Secondly, we normalized the data by dividing the difference between the relative sentiment numbers and the weekly sentiment mean (window size of mean = 7 days) by the standard deviation of the sentiment (window size of std = 7 days):
This procedure leads to data which is agnostic to differences in Tweet volume. As an example, standardized Bearish sentiment:
In the next step, we aggregated the data by weekdays, resulting in the average sentiment for each day of the week over the past two years.
Looks good …
Having sentiments plotted against each other looks very interesting. Comparing positive and negative sentiment looks as if positive is rising over the week with a positive peak on Friday while negative shows an extreme peak on Saturday.
Bullish and bearish sentiments seem to follow each throughout the week while bullish sentiment seems to dominate on the weekends.
… can be misleading.
Can we regard these trends as meaningful? To answer this we plotted each data point and sorted it by weekday. It shows that there are differences between days of the week. However, it also shows that the deviation from the mean of each day is substantial:
Implications and further work
While plotting the average standardized values of various sentiments for each day of the week would appear to give some interesting results, plotting the deviations of the data from the means shows that perceived patterns probably don’t have statistical significance.
This is not to discount the method. Further work could include aggregating sentiment over shorter or longer periods (is the market more or less bullish at the end of the month?), or by geography (which country is most bearish?). Alternative data can often contain clear periodicity that is not found in the price alone and can be used to discover market cycles that are otherwise hidden. Further work could also be done to understand how to leverage these cycles, and ultimately use them as a tool for generating alpha.