1801105 - Estimate user LTV for Fenix

Status: NEW

(Growth :: Modeling, task)

Description

[Frank Bertsch [:frank]]

This should be based on the states in the DOU models in [0] and [1].

The request is that we get LTV for:

• Organic users
• Paid users from the current MGF
• Paid users who retain from the current MGF
• Each campaign running in the MGF

If the transition matrix is singular, we may need to add epsilon and make it non-singular. If LTV doesn't make sense then we can look at the cumulative value MarkovChain.CV at e.g. a two-year period.

[0] DOU model in Colab w/ 27 inactive states
[1] DOU model in jupyter w/ 60 inactive states

Comment 1

[skahmann]

Colab notebook for Android DOU x LTV (ad clicks per client calculated on August and September campaign data)

Comments from this analysis:

• the P matrix is mostly zeros, <3% of entries are nonzero, which results in the singular matrix mentioned above
• after adding epsilon to the matrix, the LTV calculation returns positive results for all but one of the active states, and mostly negative results for the inactive states
• the CV returns positive values, this may be a better option than the direct LTV calculation