0
$\begingroup$

Long story short, I'm trying to predict how likely it is for a content creator to release new content or when they are most likely to do so (and possibly how this changes over time). My problem is that I can't find a good form of analysis/model for this type of data.

The data: the dates that thousands of different chapters from manga series (basically Japanese comics) have been translated into English. Many of these series have relatively "regular" release schedules--every week, every month, etc. Many are more sporadic, go on hiatuses, get a bunch of chapters all at once, etc.

What I've thought about doing:

  • Poisson/negative binomial distributions: most recently, I thought I could model the times between each update with a poisson process or negative binomial distribution, but I fear that the data would be way too underdispersed for that. Also, the same-periodic nature of many of these manga series violate a lot of the underlying assumptions.

  • Time-series analysis: This seems relatively logical to me, but I can't find any analyses that cover data like mine. Each data point is just a date though, so the only possible values of a dependent variable with time as a predictor would be "there was an update" or "there wasn't an update," which seems like it wouldn't fit well. Maybe I could analyze some sort of moving frequency window?

  • Periodic analyses: I've Googled for hours, reading papers and tutorials about analyses of periodic data, but I haven't found anything that feels like it could capture this data well. Both for the same reason that time-series analyses seems odd, and because the data isn't always periodic, and I'd maybe want something that makes "vaguer" assumptions.

Can anyone point me to any good resources for this, or even just give me terms I could search for? Or just helpful suggestions?

$\endgroup$
1
$\begingroup$

You can calculate difference in days between releases (time between each update) and then use exponential smoothing / arima models / plain average and any other time series technique. Average works according to Central Limit Theorem without assumptions of a distribution, but with an assumption that the process does not change in time. Croston model can be applied directly but I doubt it would be good. Maybe Holt-Winters exponential smoothing on days between releases will give good results. The best is to choose the model separately for each manga based on cross validation or plain validation (reduces time and complexity of code for time series).

You can also try calculating only difference of work days. On Kaggle the winning solution of predicting sales for a top-5 retailer in the US had a very complicated manipulation of data based on holidays. (Walmart winning solution) Your days will be dependent on number of days in a month.

And like on Kaggle you can also use simple models for ensembling multiple models or even as a standalone models before choosing based on validation.

The Kaggle solution also contains lots of time-series models that you can try on "days between" data.

In Croston the data is called "intermittent demand" but it can have multiple "issues" at one day where you can have only one.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.