I've watched a truly embarassing amount of anime this year. As a result, as any anime nerd can attest, I've spent a lot of time on MyAnimeList (often abbreviated as MAL), a popular anime forum where shows are rated, categorized, and recommended. It's a kind of IMDB for anime.
After some time scrolling through MAL, I noticed that sequels (seasons 2+) of a show seemed to be ranked higher than the first season. I wondered why this would be the case, since (in my experience) later seasons can go downhill just as often. So I decided to test my theory with science!
I assembled a dataset of 30 top-rated shows1 on MAL which have at least two fully completed seasons, together with their ratings for both seasons. I found that out of these 30 shows, only 2/30 have a lower-rated second season compared to their first season 2. On average, the second season is rated 0.27 ± 0.16 points (or 3.3% higher). This suggests that there is a real rating inflation for higher seasons (what I call "sequel rating inflation" going forward).
This result is statistically significant with p<0.01, meaning it's a real effect.
Why?
For why this is the case, my theory is that people who don't like a show will stop watching it after the first season (and give that season a low rating). Therefore, the people who watch subsequent seasons are more likely to be fans of that show. I don't have any data to support this conclusion, though.
Making an Adjustment
So what if you want to know if a second season of a show is really better than the first, or it's just due to "sequel rating inflation"? Simple - just deduct 0.27 from that show's Season 2 rating.
Using this adjustment, our results seem to make sense in the chart below based on anecdotal reviews on the anime subreddit. Shows with a second season that's considered better than the first (e.g. March Comes In Like a Lion, Kingdom) retain their higher rank. Meanwhile, shows with lackluster second seasons (e.g. Spice and Wolf, Bakuman) now have their second season rated lower than their first season. So the results correspond with our intuition.
Closing Thoughts
I've uploaded the dataset and notebook onto GitHub if anyone wants to try this experiment with a different dataset. This analysis might look different if you use a bigger, or different, list of shows. With a dataset of this size, it might be easy to skew the results by adding one show with a very dramatic rating drop (e.g. The Promised Neverland). Let me know what you think by reaching out on Twitter.
