Re: Time Series Analysis on Distributed Merits in the forum (daily, weekly, monthly)
Sure, I will. This is my original intention.
What I have posted are only descriptive ones. More things to come later.
Something like this one:
I am not sure, but I guess it probably come from effects of two factors:
1) New merit sources in June. Not sure, because I can not check.
via the link: https://bitcointalk.org/index.php?action=merit;stats=sources, I can get only statistics on merit sources for today, not the past ones. I meant we might get something interesting if be able to retrieve time-series data of merit sources, and a total merit generation of up to 22045 sMerit per 30 days of those sources per day. Maybe weekly data is enough because @Theymos probably have not spent too much time to check and add new merit sources per day.
I don't familiar with this, and can not retrieve the data, but you probably can do this with your skills (I guess).
2) Something behind the scene, from the dark side of the forum. Something like suddenly rocket of autobanned accounts (high-ranked ones, of course).
Anyway, thanks for your ideas, from now on, I will keep tracking these two figures (total merit sources, and their total smerit generated per month).
I will update the OP with truncated daily data analysis.
What I got:
IQR = 788 525 = 263;
1.5*IQR = 1.5*263 = 395
Potential outliers are days with total merits above 1183 or below 130. The mentioned day (153th in the dataset) with 1138 merits in total is likely a potential outlier, very close to 1183.
- Potential above outliers: Q3 +1.5*IQR = 788 +395 = 1183
- Potential below outliers: Q1 1.5*IQR = 525 395 = 130
Days with figures outside the interquartile range (from Q1 to Q3) should be considered as potential outliers, and days with figures outside the Q3 + 1.5*IQR and Q1 - 1.5*IQR are highly potential outliers.
Moreover, the median of daily merits from truncated dataset is 627, not change too much than the rest two dataset at 653 and 650 for full, and partially truncated ones, respectively. Median shows magic meanings, again.
I surely will do it regularly, because it is my purpose to start the topic.
Highly appreciate your help by giving me daily data (updated on weekly basis is enough).
By the way, I would clarify that the weekly statistics in my analysis are different than yours.
Why?
I used you daily datapoint, then converted them to weekly one one the structure of real calendar.
It means that the first week in the data set (from 24th to 28th of January) contains only five days, Wednesday to Sunday.
I will keep using the approach for my weekly analysis.
What I have posted are only descriptive ones. More things to come later.
Something like this one:
Can you replace the linear "Fitted values" by something that fits all values better? I think a logarithmic fit could work.
I am not sure, but I guess it probably come from effects of two factors:
1) New merit sources in June. Not sure, because I can not check.
via the link: https://bitcointalk.org/index.php?action=merit;stats=sources, I can get only statistics on merit sources for today, not the past ones. I meant we might get something interesting if be able to retrieve time-series data of merit sources, and a total merit generation of up to 22045 sMerit per 30 days of those sources per day. Maybe weekly data is enough because @Theymos probably have not spent too much time to check and add new merit sources per day.
I don't familiar with this, and can not retrieve the data, but you probably can do this with your skills (I guess).
2) Something behind the scene, from the dark side of the forum. Something like suddenly rocket of autobanned accounts (high-ranked ones, of course).
Anyway, thanks for your ideas, from now on, I will keep tracking these two figures (total merit sources, and their total smerit generated per month).
Quote
Any idea what happened on Day 153: 1138 Merit? I noticed that already when I first produced the data.
Now, let's take a look at truncated dataset (I truncated all days before 19th February, the same approach as of weekly analysis).I will update the OP with truncated daily data analysis.
What I got:
1.5*IQR = 1.5*263 = 395
Potential outliers are days with total merits above 1183 or below 130. The mentioned day (153th in the dataset) with 1138 merits in total is likely a potential outlier, very close to 1183.
- Potential above outliers: Q3 +1.5*IQR = 788 +395 = 1183
- Potential below outliers: Q1 1.5*IQR = 525 395 = 130
Days with figures outside the interquartile range (from Q1 to Q3) should be considered as potential outliers, and days with figures outside the Q3 + 1.5*IQR and Q1 - 1.5*IQR are highly potential outliers.
Moreover, the median of daily merits from truncated dataset is 627, not change too much than the rest two dataset at 653 and 650 for full, and partially truncated ones, respectively. Median shows magic meanings, again.
I surely will do it regularly, because it is my purpose to start the topic.
Highly appreciate your help by giving me daily data (updated on weekly basis is enough).
Quote
Will you do regular updates? I can provide input data weekly, it's not much work:
By the way, I would clarify that the weekly statistics in my analysis are different than yours.
Why?
I used you daily datapoint, then converted them to weekly one one the structure of real calendar.
It means that the first week in the data set (from 24th to 28th of January) contains only five days, Wednesday to Sunday.
I will keep using the approach for my weekly analysis.