The best algorithms to detect continuing decrease pattern on conversion data

user_1921025 0
Viewed 125

I'm tring to use an algorithm for finding and ranking continuing decreases in my data. The base condition:

The breakpoint location (blue arrow) is unknown but bigger than x (can be on t= 4 or in t=9 but can't be smaller than t=2 for example )

Part B (after the breakpoint) has smaller AVG than Part A (before the breakpoint) significantly given the data variance.

The larger the ratio between part A and part B, the higher the score.

Need suggestions for the best practice.

enter image description here

UPDATE: I spit the X axis to last 2 days and 10 days before those days. and used SUM((x_last_2_days -x_avg last_10_days)/std_over_all )

So I got the following patterns

Rank 2 - fine enter image description here

Rank 3 - controversial enter image description here

Rank 6 - bad enter image description here

Rank 7 - good enter image description here

Rank 8 - bad enter image description here

How to ignore the bad pattern or improve the algo.

anomaly-detection math algorithm
1 Answers

It looks like you are dealing with a noisy time series. Your problem as formulated is known step detection or edge detection. Your approach to solve it is listed in the wiki there (look for sliding window). Because the data is noisy, you do have to employ statistical methods to detect features in the noisy data to differentiate a change in the pattern from pure noise. Concrete implementation would require some analysis and possibly simulations.

In general, analysis of time series and signal processing involves de-trending, de-seasoning, edge detection, so that all that remains is pure white noise. Tools used are Fourier transform, autocorrelation, etc.

user_105496620
Related