#fanalytics

Sometimes you look at a chart and it just makes sense, and for me it was this wonderful visualisation of the Sherlock fandom by @fffinnagain​ that came across my dash the other day. Full disclaimer - I didn’t even read their full post, but I think this is one way of looking at the data that speaks multitudes for itself. 

Take a look here and tell me what you see!

I’ve divided the “waves” here into a few categories - “On Air”, when ATLA was airing for the first time, “Pre-LOK”, the time between ATLA and LOK, the “LOK era” which is self-explanatory, and then every year individually after that. A creator is assigned into a wave based on the date of their earliest fandom work (due to the way AO3 works, this isn’t always the date they published their first work, but rather the date this work was last updated). 

Once the creators are assigned to their groups, the colors are used to track any additional works they produce in the fandom. Such works are counted up per month, which explains the small local variations in these waves. I’ve marked a few key dates on the graph as well - one or two of which might explain the 2020 boom ;)

The ATLA Renaisance of 2020 brought almost 10x the amount of new creators to the fandom as any other year! How many of these fanwork creators will be around in a couple years? Well, judging from previous waves, it does seem that contributions from creators who entered a fandom die off to a trickle come the next year. The bunch that entered during the LOK era seem to still be contributing consistently however, and there’s a definite resurgence of the OG fans coming back to the fandom, as you can see from the little blue-purple bump in the year 2020. 

We’ll have to wait and see if the new fans are here to stay - but for now we can enjoy the wonderful content they produce!

Let’s take a look at the next-gen Avatar fandom, too. Overall there are far less works in the Legend of Korra tag, although its release on USA Netflix has given it a popularity boost as well - one that’s even bigger than the one it got from the landmark series finale.

I’m interested in eventually examining the crossover of creators between these two fandoms, to answer questions such as: how many creators participate in both fandoms, and of those, what’s the fandom of origin and the lag between the crossover? 

Anyway, for now, tag urself I’m a 2019 ATLA baby, and I still can’t believe I’ve never written a stand-alone for LOK (side-eyes my fanfic drafts folder).

AO3 statistics for the Legend of Korra, 2020

[image: Two screenshots from Google Data Studio dashboard, showing statistics from the LOK fandom on AO3 in 2020. Headers are: Total new or updated works, Rating, Category, Characters, Relationships, and Tags. Totals are shown in pie charts or bar graphs, displayed along with timeseries data smoothed with a 7-day rolling average.

A noticeable increase in the number of works happens shortly after the release of LOK on USA Netflix in mid-August. This is a great fandom for wlw, with almost half the works being F/F and Kyalin bursting onto the scene in the fall. Endgame ships remained strong, with Korrasami ~3x as popular as any other ship, but good support for Bopal and Kataang (though probably as background). One mlm ship, Mako x Prince Wu, is present in the top 10, and seems to have maintained a relatively small yet steady interest since the US Netflix release. 

6/10 of the most popular characters are female, represented proportionally in the number of total works. 

Interact with the data at https://datastudio.google.com/s/j2oYkqzahFA, where you can use the mouse to hover over selected points or select and deselect timeseries data of interest, and hopefully Google had the foresight to make this all compatible with screen readers :) ]

Played around with Google Data Studio today, so have some fandom stats 2020! Explore the data interactively by clicking the link.

Verdict on Data Studio: very fast and easy, decent style, but lacks functionality (I had to do the rolling average on the dataframe and then export to csv and upload to Google Sheets. Also, it somehow lacks the usual Plotly/Tableau capabilities to zoom and select/deselect traces at will on any plot). I will definitely use it again for fan-related stuff though, since the hosting is free and basic is good enough :)

I asked this question about the AO3 ATLA dataset a few months back, tackling the problem with multiple linear regression. Since then I fixed a few bugs in my scraper and got an updated version of the dataset (and wow has the fandom taken off last year!), making it now more than 10,000 fics.

I also applied a decision-tree based algorithm, XGBoost, which generally gives better performance than pure regression on small-to-medium datasets. Like last time, the value I wanted to predict was the kudos-to-hits ratio.

Let’s see how the results stack up!

According to multiple linear regression, these are the most important features that contribute (either positively or negatively) to a one-shot’s popularity. The overall largest feature is Year, which didn’t fit on the scale of this chart :) After the year comes some categories and then ships: Zukka, Maiko, and Kataang. As a Maiko shipper myself, I’m got the feeling that this is actually a factor that makes your fics less popular... there’s a bunch of readers who will not even click on a fic that includes that ship. Same with Kataang, if you’re a hardcore Zutara shipper.

Not all of these factors match up with my intuition. While Modern AUs are certainly popular and Found Family is a core trope of the franchise, does having Pakku in your fic really help or hurt? Or the (rarepair?) of Katara and Suki?

XGBoost selects somewhat different features as important. It also fit the test data better than multiple linear regression.

A lot of the top features are the same, but with M/M switched with F/M. Another interesting feature which totally confirms my gut feelings is that of Author_wave, defined here in a Creator Waves analysis. Basically, this is the era at which a fic author entered the fandom. So it stands to reason that well-established authors will continue to write popular fics.

By the way, here’s how the two algorithms managed to predict the test set data. The tighter distribution of XGBoost means it did an overall better job, but there’s still a bunch of fics it didn’t manage to predict well. XGBoost tends to do very well with classification (for example, popular or non-popular) so I play around with that.

November 2020 was one to remember -and for fans of My Hero Academia, it wasn’t necessarily for the US elections. The long-awaited confirmation of the fan-favorite theory concerning the villain Dabi’s true identity arrived in the form of Chapter 290. 

So how was this reflected in what creators put on the archive? On average, there was a 10% decrease in the use of the two similar tags Dabi is a Todoroki and Dabi is Todoroki Touya.

This isn’t immediately obvious when looking at the day-by-day numbers of fics with the Dabi character tags updating or being published (grey line). However, applying a 30-day rolling window makes the trend clear (orange line). 

Behold, the power of maps x web app! The live version will be released when I finish analysis, so that you can explore the data yourself!

Above: Map of the world with locations mentioned in modern ATLA AUs indicated by: color intensity (mentions of countries and US states) or bubble (mentions of cities), over time from 2012 - 2020. Values accumulate as the year slider moves.

Below: Map of the ATLA world, with locations mentioned in canonverse works indicated by bubbles (until counts grow above 100, in which case their size is reduced by a factor of five).

What can I say, I be punkin’

Here’s my fanalytic fm

Aaand here’s my c.ai wrapped (har har har…uhh yeaahh I have no life…)