Since the fluctuation of the indicator can be divided into normal and abnormal, and according to the duration, it can be divided into temporary fluctuation, periodic fluctuation and continuous fluctuation. Therefore, it is necessary to clarify whether the fluctuation is really abnormal. Operation, R&D, operation and maintenance and other parties confirm the authenticity of the data, and then observe how long the data fluctuations have lasted to determine whether the fluctuations are within a reasonable range, compared with yesterday and last week, how much has changed year-on-year and month-on-month, etc. The reasons that may be found in this step are: abnormal server, background statistics error, abnormal report data, inconsistent index calculation caliber, etc., or it may be caused by external malicious behavior, which needs to be carefully checked according to the subsequent steps.
"Do split", dismantle indicators and business processes in multiple dimensions, and narrow the scope of investigation In this step, pay attention to the "Simpson's paradox", that is, in some cases, analyzing the data as a whole and analyzing the parts of the data will lead to opposite conclusions. 1) Dismantling from indicators : An indicator can generally be calculated from other indicators, and the fluctuation of an indicator is the result of the combined mobile number list action of several factors. If you want to narrow down the scope of the problem, you can try to subdivide the original index as much as possible, and continuously disassemble it from coarse to fine to find out all the factors that may affect the fluctuation of the index, and get a tree-like structure.
For example, "DAU = new users + retention of old users + return of lost users", new users have source channels, old users have retention channels, lost users have natural return and intervention return, and the common factors of users include software version, region, equipment , active time period, etc. In addition, splitting is not only by addition, but also needs to be split according to specific indicators, such as some composite indicators: "GMV = new user x conversion rate x new user customer unit price + old user x conversion rate x old user customer unit price". 2) Dismantling from the business process : in the whole process of a user from registration to entering the homepage and then to loss or retention, any step in it may be the "culprit" of the decline in indicators.