How to do a monthly or yearly analysis in RFM?

[franzoni315]

In my company we sell yearly and monthly subscriptions with possibilities of cross-sell and upsell. How can I use the rfm_summary with monthly or yearly periods? When I try to set time_unit="M", I get this error:

ValueError: Unit M is not supported. Only unambiguous timedelta values durations are supported. Allowed units are 'W', 'D', 'h', 'm', 's', 'ms', 'us', 'ns'

[ColtAllen]

Hey @franzoni315,

Use the time_unit and time_scaler parameters for this:

rfm_summary(data,
            customer_id_col,
            datetime_col,
            time_unit="W", # "M" & "Y" are unsupported TimeDelta units in pandas,
            time_scaler=4.33, # so Weeks must be rescaled into Months
            )

[franzoni315]

It worked, thanks very much, @ColtAllen !

[franzoni315]

@ColtAllen I actually discovered that this parameter is not available for the expected_customer_lifetime_value function. I am trying to adapt the CLV quickstart notebook for a yearly based business model, so for the rfm_summary I used the parameters time_unit="W", time_scaler=55, but I can't do the same for the expected_customer_lifetime_value function. I tried to set

clv_estimate = gg.expected_customer_lifetime_value(
    transaction_model=bgm,
    data=rfm_data,
    future_t=55,  # year
    discount_rate=0.01,  # monthly discount rate ~ 12.7% annually
    time_unit="W")

but I am not sure if this is correct, and also it takes too long.

[ColtAllen]

@franzoni315 sorry for not seeing this sooner; I get a lot of emails related to pymc-marketing notifications.

I actually discovered that this parameter is not available for the expected_customer_lifetime_value function.

It's not used in this function because rfm_summary aggregates data for model fitting, not predictions.

clv_estimate = gg.expected_customer_lifetime_value(
    transaction_model=bgm,
    data=rfm_data,
    future_t=55,  # year
    discount_rate=0.01,  # monthly discount rate ~ 12.7% annually
    time_unit="W")

but I am not sure if this is correct, and also it takes too long.

Set time_unit='M' and future_t to the number of months in the future you want predicted. "Years" is an unsupported time unit, but equivalent to future_t=12*N_years. This will also require the transaction model to be trained on monthly data.

Also, is this model only being used for cross-sells and upsells? If you're trying to predict if a customer will renew a subscription, ShiftedBetaGeoModelIndividual is the intended model.

However, for large customer populations ShiftedBetaGeoModelIndividual may take too long since it'll generate unique posteriors for each customer, and it can't currently be used with the expected_customer_lifetime_value function. This could be the proper motivation for us to create the population equivalent of this model and add discounted CLV support.