Consider Last-Year Trends into Expectation Range at Performance-Charts
complete
Andreas Wagner
complete
Jan Hempel
thanks, this seems reasonable now.
The boundaries go up in november and down in december even though the previous 5 weeks alone would indicate something else.
Well done!
Jan Hempel
I cannot judge it based on your answer. I think not.
I am not asking to just usr more data for the confidence interval, but use it also more intelligent, season based.
So comparing first november week with past five weeks plus a seasonal compnent that takes into account the usual trend in first november week.
Andreas Wagner
Jan Hempel: Hi Jan, sorry but in order to identify seasons (which you refer to as more intelligent usage) we need to use more data to first identify these seasons (as we are not going to hardcode seasons) - and that was the whole purpose of the task - in other words the new lines in the charts represent these seasonal fluctuations and I wanted to know if they look like expected or not :-)
Andreas Wagner
Jan Hempel: Hi we found another approach which does the following:
1.We take all the data from last year remove extreme outliers and smooth the data based on seasonal effect which we extract from all the data we have for a given tenant. (Whittaker-Eilers method)
2.We then calculate the confidence intervals and show them together with the data of the current year (time-frame of interest)
- By doing so we ignore short term outliers if they are expected via prior data (as wished)
see the following chart which shows some sample data for search CTR
Andreas Wagner
Hi Jan Hempel,
we included the confidence bands based on the data from the last 1.5 years and included it in the following screenshots (green line represents the lower bound while the orange line represents the upper bound. Is this what you would expect to see?
Andreas Wagner
in progress
Jan Hempel
Yes, either you could provide a separate confidence interval or try to incorparate both short term and annual trend into one
Andreas Wagner
under review
Andreas Wagner
Hi Jan Hempel,
ok so if I understood correctly another confidence-band based on the 52 weeks or 365 days of the previous year would do the trick.
Please let me know what you prefer - FYI the current confidence bands are always calculated based on weekdays, as the behavior is very different them.
Andreas Wagner
Hi Jan and thank you for the feedback.
Can you please elaborate on what you mean by "considering the trends of the year before" ?
What does this entail in detail? What should be compared?
Jan Hempel
Hi Andreas Wagner,
currently, when you only consider the past weeks, you ignore annual trends.
Therefore, the expectation range during high or low peak seasons is quite useless at the moment. during black friday week, the kpis are all above the range, and then afterwards they are all below the range.
Including annual trends and combining them in a weighted fashion with recent trends might increase the explanatory power of the expectation range.