Cohort split guidelines (CLV)

[lixgl]

Hey team,

While browsing through original authors notes on BG/NBD model and was wondering how frequent the cohort splits should be and if there are general rules of thumbs there

While I saw some suggestions of splitting data strictly by all purchasers who made their first transaction on a specific day there were other suggestions of ignoring the cohorts entirely and feeding the model with all available customers transaction data (so it could take into the account "older" buyers too)

[ColtAllen]

Are you interested in constructing the cohort table in the spreadsheet the research note describes? This plotting function summarizes said table into a time series plot.

was wondering how frequent the cohort splits should be and if there are general rules of thumbs there

If I understand correctly, this is a choice of time period aggregation and selection; it depends on your use case. Some examples:

• If the model is to be used for reporting at a monthly level, aggregating in clv.rfm_summary for monthly predictions could make sense.
• If you have strong weekday seasonality, aggregating by week would result in a more "behaved" dataset for model fitting.