According to statistics kept by the Eve-Offline status monitor, the average number of players online:
Since Launch: 37,000
Last 5 years: 36,000
Last year: 33,000
Last 6 months: 32,000
Last 3 months: 27,000
Last month: 24,000
This despite the pandemic fueling new interest in MMOs. New account creation seems to be up. The problem is retention. As somebody who recently returned to the game, various chat channels that should be expected to have chatter are dead.
Whatever CCPs plan is for 2021, they better toss it out and focus on building player communities. The only thing keeping these number from being apocolyptic is that they cover the holiday season, generally a down time for EvE (though I remember a couple sieges on Christmas and New Years in the 2000s).
Well the main problem I see with that hypothesis is that it’s flat wrong. The drops from downtimes are going to make up the same percentage of the average no matter what time scale. Also the three month sample size should be a large enough that the margin of error should be 10% or less. This change is outside of the margin of error.
There appears to be a serious problem with player participation in this game.
CCP is ignoring building player communities. The phone MMOs like Lords Mobile (that steals heavily on gameplay ideas and lessons from EVE) put the player community first - which drives a lot of ridiculous spending.
CCP took a nerf back to ratting in populated systems and mining. Normally these are actions that are done with other people. This destroys the community interaction. Further, they are shredding players sense of place, both by sparking a huge 0.0 war that is litterally destroying places and shredding places through the trig invasion.
From what I’ve seen of the current war, it differse drastically from what happend in 2008. Players are fighting to protect their community. Because of the resources shortage this is causing what might be termed a grim war. People are liquidating the assets on their account to figh the war, then quitting the game.
No, it’s absolutely correct. That average takes all data points and thus the larger the window, the more “maximum values” are used. Just look at the one year graph. Does it at all match the numbers you claim? There is no way that the PCU average in December was 24k.
Um, we are talking about the average of a cyclic time series. That is something that needs to happen a lot in science and engineering. If what you say is true, you should easily be able to find and cite documentation explaining the issue.
The longer average includes more maximum datapoints, which are divided by a larger number. It also includes more cyclic lows.
Login numbers are down. Multiple metrics - including counts of login numbers - show that login numbers are down. This is a problem that needs to get addressed.
I know. This is simple math. I am not sure why you aren’t getting it.
Well, if you didn’t get my explanation before, let me try one last time.
This is a made-up dataset using a sine function:
The data here is the complete data set and as expected the average is around 0 as you would expect (0.028 to be precise). You can take the average of this it is does reflect the underlying process as you have sampled it fairly.
If however you compress your sampling and start combining your data points using a maximum function, you will completely skew the average. Taking these exact same data points and reducing them by 1/3 by taking the maximum of each three points, you get something that looks like this:
The average is 0.595 now and completely wrong. The underlying data is the same as the previous curve, just combined via a maximum function, but now the average of these max projection points does not reflect the original data at all.
This is exactly what Eve Offline is is doing. If you look at the datapoints, they have the full time stamp and you can see that for the time interval for the data is about 30 minutes for the one week charts (so the max player counts in a 30 minute window), and over a day by the time you expand to the one year chart (so the data points are the max players over the course of a full day or more). Since @Chribba is tracking the maximum player count during these time windows, it makes sense for him to choose the maximum player count value to display when changing the sampling window size like that, but as a side-effect people keep missing for years now, that means the averages are not comparable between different sampling windows. You only can compare the average data when sampled with the same time window.
Look, I am going to draw the 24k average you claim for the last month on the 1 year chart in green:
How does that match the data? My best eye-balled average (in red) is somewhere around 31k, which ok, does seem to be down slightly from the 1 year average, but still well in the noise and above the January/February 2020 average we started the year with.
The 24k number is simply because more of the data points reflect quiet times of day revealed by taking smaller sampling window sizes. When you look at a year, these datapoints are combined and only the maximum is used in the average.
Overall activity, at least judged by this metric isn’t down. Maybe specific areas of the game are, or other metrics might show something different as the PCU isn’t the greatest metric, but that is the point. This metric is dubious at best, and doesn’t even show what you claim.
Thank you for that explanation. I’m no math genius myself and tbh honestly also never put super much effort into the average value, and with the amount of data points I had to do some pre-calculations to speed things up (legacy code), so indeed a number of charts use max() rather than overall average of all data points (plus weeks/month/years are pre-cached on intervals, whereas 36h for example is pulled in real-time.
I am happy to assist to pull specific numbers, using whatever methods you guys want to give a better view if wanted (short of giving all raw data points, but I can do that using calculations you ask for).
It’s fine as it is. The data and average are correct, just not comparable creating the confusion of the OP. You are reporting the Peak Concurrent User value of each time window so it makes sense to use max() to get the peak value.
I guess one improvement you could do to make it comparable would be to calculate the average line with only the daily maximum for every day present, not the average of each data point. So no matter if it is the 1 week view or the 5 year plot, just take the PCU of each day of the range for calculating the period average (or maybe a function could discard the non-max points for each day, just keeping one point, the peak value of the day, for the averaging?). The average line wouldn’t go through the middle of the data points now, but would be comparable across the views and be better suited for what people seem to like to use that average you calculate for.
But if you don’t have that data already in a table, I am not sure it is worth the effort or extra computation time.
Thanks for such a longstanding and useful resource though! It is great to be able to visualize this information across such a long time.
Perdro that would be a great argument, if I had used two week or less data. The one month data that I used is a daily high and low which is on the same time scale as the 3m, 6m, 1y 5y and All. I see your point, I simply did not make the mistake you are accusing me of.
Even the 3 month plot seems to have about 4 hour windows for each data point. You can clearly still see the daily cycles, all of which are being used to create that average line:
See how the data point is labeled Dec 27, 2020, 10:30? The next one is labeled 13:00. That data point is the max value of all the data Chribba collected in that time. That means like 4-5 time points per day, so you basically are averaging each time zone, including the much quieter ones, not averaging the daily max PCU as you do in the longer charts like 1y or more.
The larger the time window of the chart, the larger the time window of the data points, and since it uses max(), the average will trend up as the peak times of days dominate the the quiet times of day. You can see this by eye as the daily cycles disappear as the time window increases, but you can also hover over two adjacent data points to see the size of time each data point represents.
Compare within a single chart all you want - like use the 1y or 5y chart and draw some lines - but it should be obvious that the average a bunch of Peak Concurrent User (PCU) points will depend on window of time you sample. If you average the PCU every minute over a day, you will get a number roughly half way between the busiest time of day and the quietest. If you take a time window of the whole day instead of a minute, the data in the average will be equal to the busiest time of the day and be much higher, because you are taking the max value in that window.
This completely explains your observation in the OP - it is an artefact of the sampling window. Looking at the 1y chart there is no evidence of the massive decline you claim from using the average lines on different charts.
There is nothing like this in the Eve APIs. You can get some analogues for activity though that are available on Adam4Eve (and on Dotlan in a less granular way), like gate jumps or NPC kills: