Prophet to detect anomaly.

[jeesim2]

Hello.

Prophet seems really simple and strong at forecast as I tested for several days.

Now I wonder I can use prophet to anomaly detection.

Basic idea is... first, fit model with data from a year ago to several days ago. second, run model.forecase() for recent and near future days. third, when actual data collected, compare forecast to it. If actual data far from forcasted I will count it anomal.

Problem is I have no idea about third step. I just read some articles like this. https://www.google.co.kr/url?sa=t&source=web&rct=j&url=https://blog.statsbot.co/time-series-anomaly-detection-algorithms-1cef5519aef2&ved=2ahUKEwiXz4vGkoXcAhUIm5QKHY1kA7UQjjgwD3oECAQQAQ&usg=AOvVaw10vigSY6syGsS-h9KjM6E3

Is this considerable approach to do anomaly detection?

[IronistM]

This is something that I have used in the following context, get event counts from Google Analytics and compare it to the expected. If the difference ie higher than a threshold then mark it as a review point. You can check it out here googleAnalyticsProphetR.

The tricky part is not to get overfit and/or how to define the threshold for marking something for review.

[jeesim2]

@IronistM googleAnalyticsProphetR can be a good example for me.

BTW, I see you use tweeter's AnomalyDetectionTs function. Did you have tried any other function to pick anomalies?

It will be good If there exists AnomalyDetectionTs alternative for python.

[IronistM]

You can check out nicolasmller's pyculiarity and Marcnuth's ports.

[Diyago]

@jeesim2 Take look at https://towardsdatascience.com/anomaly-detection-time-series-4c661f6f165f

[abhimanyu3-zz]

@Diyago How the algorithm will perform if we will give minute by minute data instead of day or monthly data as there is lots of noise in minute by minute data stamp