## Channel Title: Avanti Gurukul ## No. of Views: 12920154 ## No. of Subscribers: 190973 ## No. of Videos: 360
Reading Lifetime Subscriber Logs downloaded from YouTube Analytics (Max 500 rows per CSV at once)
Cleaning subscriber logs
August 16, 2018
Connect to Avanti Gurukul channel via YouTube API
Plotting all variables vs date
Note that average % viewed can be over 100% if the user runs the video multiple times. Source : https://www.quora.com/On-the-YouTube-analytic-report-why-is-my-average-view-duration-in-certain-countries-above-the-actual-video-duration-and-the-%7Faverage-percentage-viewed-above-100
That's a useful graph. Let's clean it further and check
Let's try and model daily subscriber counts
Clean daily subscriber graph
Filtering dates from Aug-2016 to Aug-2018
Checking for a linear trend
Residuals from a linear model fit
For a good model, residuals will appear random. But here, they still increase as the date increases. 
Log Transformation
Residuals from log model fit
Log transformation removed the increasing trend of residuals !
Seasonality - Weekly
Weekly seasonality is not strong for this dataset. However, there is moderate day-of-week seasonality since May-2018 with Wednesday a clear peak. 
Seasonality - Monthly
Monthly seasonality is fairly strong in the last 2 years 
It seems we've passed our peak. Winter is coming !
Adding monthly seasonality to the log model
Residuals from log model with monthly variations
Barring irregular peaks, eg.,March-2018, looks much better 
Preparing the test set using data from Aug-1 to Aug-13, 2018
Making predictions on test set
Predictions look okay for the first 13 days of Aug-2018 
Create a prediction set for the rest of 2018
Making predictions
So where will we be at the end of 2018 ?
Summing up daily subscribers from prediction
On 31st Dec-2018, our subscriber count will be :
best_case
## [1] 342282
expected
## [1] 299825
worst_case
## [1] 268813