Hello all!
I am working on a time series model by using FB prophet. One of the issues I have is that in certain days, the model is consistently over or under forecasting. General diagnosis I have for these over or under estimations indicate potential outliers on those dates. However, on those dates, there are reasons for variation/outliers but the magnitude may not be taken into consideration accordingly by the model. What I mean by this is that we do have certain events happening on these dates that we do expect it to go up & down based on the changes in some of the variables. It is adjusting to the variable to update the yhat but I can not seem to figure out the reason in some days. Lets say every 4th Thursday of the month it overestimates but in other 3 it is within the reasonable error margin.
Any ideas or suggestions as to where I should focus on to figure out this very particular issue?
I have looked into many potential external factors that may cause these changes but so far have not been able to solve the issue with any additional variable hence I am now looking into the y itself to see if there are any obvious things that I am missing.
Looking forward to brainstorming with you all.
Outliers are in the eye of the beholder since you set the criteria. I have attached some resources you might consider, esp
Cont. Esp these two: the second is more general, the first one is by the time series expert on outliers, available from the z-library. Good luck to you. David Booth
Yes as per the definition, an outlier in one analysis may not be an outlier in another analysis. Depending upon the nature of the problem for which we are looking for rightly said by David Eugene Boot Outliers are in the eye of the beholder since you set the criteria, and further sources can be checked.
- Badges
- Science topic
- Similar topics
- Statistics
- Time Series