I've been reading around your blogs and guidelines on anomaly detection and contribution analysis, and this looks very powerful. I'm keen to make sure we interpret this correctly and can explain the methodology to others. I was wondering if you could post or email some information around the stats side behind this.
If I understand correctly Anomaly detection is using Holts Winters Multiplicative, Holts Winters Additive and Holts Trend Corrected, is that right? If so could you please advise a source that explains these in a straight forward way. Also could you explain how Adobe applied this, especially. E.g is the seasonal factor a daily factor in this case ?
Likewise on contribution analysis, I believe this uses Pearsons Residual, is that correct ? Again is there a straight forward source I can read around? How does Adobe apply this ? In particular the contribution score , how should this be read, in the examples touched on how does a dimension item with only 1 bounce have a high contribution score? I can see it may have a significant move, from 0 to 1 but wouldn't the low number of bounces reduce its overall contribution score.
Very early days in looking at this so I may be interpreting completely wrong, any help greatly appreciated
You are right about the anomaly detection, it uses Holts Winters Multiplicative, Holts Winters Additive and Holts Trend Corrected algorithms. I am assuming you went through the whole documentation on anomaly detection and contribution analysis, if not here are the links to the same
It is difficult to reveal how Adobe applied this because that is something internal to the company but I would try to find out if there is a source to explain this in further detail for you. Feel free to ask further questions around this. Thanks!
HI Tanmay, thanks for the quick reply. I've read around all the docs but still struggling to fully grasp everything. Any docs you can share would be appreciated.
Also I ran a contribution analysis today which made no sense to me, can you help. It was showing an anomaly where bookings were less than expected. Let's say it expected 1000 but was much lower at 800. In the top items it read contribution score .45 and bookings was showing as 800 (100%). This confused me as I thought a high anomaly on a negative anomaly would show as a negative number . Also I thought the metric (bookings) would only show a maximum of the anomalous amount I. E. 200 so not sure why it's showing 800 ?
I have been looking at the Anomaly detection and the contribution analysis. Looks very cool and great tool to have, however:
1. All the help center resources are very outdated and not inline to the new UI and way it works. Any idea as of when it will be updated?
2. I also agree with the contribution analysis being very confusing, especially when it allows you to list the top 5. I constantly need to keep hiding elements in order to display new ones. Looking at an anomaly based on a dip of data i would expect to see the contribution score closer to -1 first, unless i am not getting how it works.