Objectives of the project:
- define seasonal components
- achive stationarity
- predict dynamic of new trending video release on YouTube
Before I dive into any statistical methods for predicting future data, let's take a look at the data we already have.
We may see daily dynamic of new video releases in first quartile of 2018
From here I can pick out an immediate trends
It looks like there could ve some structers behaviour in the weeks themselves, some signal from the day of week.
**Here we can see a few different weeks.
It seems there is a pattern in hourly video releases within a week: new video publushing more likely in the middle of the week
If this assumption is true I could see a strong seasonal component to the data.
Here we can see Tue, Wed, Thur, Fr are highlighted
The most striking feature here is the oscillatory nature of the autocorrelation plot. This is clear indication of non-stationarity.
Lets apply differencing that should result in a stationary series.
This looks possibly more stationary. Lets run hypothesis test to be more confident
ADF p-value: 0.00
Ok, there is a time to make a predictions for the next four weeks
The prediction for the next four week is clear, there should be a downtrend in numbers of new video release. And it is. You may check youself.
- We defined seasinal components, such as weekly releases of new video. The videos publishing within the work day mostly.
- We achived stationarity by applying differenciation.
- Prediction was made. I am betting it is correct.