search

pymc-labs / pymc-marketing

Bayesian marketing toolbox in PyMC. Media Mix (MMM), customer lifetime value (CLV), buy-till-you-die (BTYD) models and more.
https://www.pymc-marketing.io/
Apache License 2.0
648 stars 175 forks source link

Investigate survival model alternatives to existing lifetime models #277

Open ricardoV94 opened 1 year ago

ricardoV94 commented 1 year ago

Could be more flexible:

  1. Easier to integrate covariates
  2. Easier to customize (priors, likelihoods, arbitrary structure)
  3. More similar approach to MMM (have on robust model that can be customized)
larryshamalama commented 1 year ago
  1. Easier to integrate covariates

Do you mean incorporate? This is very nit-picky for now, but integration of covariates within longitudinal (especially causal) has a very specific meaning (c.f. G-Formula or G-Computation).

My understanding is that the bigger goal for CLVs is to relax our model assumptions. In biostats, these purchase times can be referred as recurrent events or, when an outcome (e.g. purchase amounts) at these times are of primary interest, irregularly observed data or even covariate-driven/outcome-dependent/etc. observation times. I like Pullenayegum and Lim (2016)'s review paper and Ryu et al. (2007)'s paper, the latter is Bayesian.

Modelling of the outcome and timings can be done via joint modelling, i.e. having dependence between models parameters. All of this can also have sequential interventions, which can add causal layer to complicate things further (or make things more interesting).

This opens up sort of a rabbit hole in longitudinal modelling that I particularly enjoy, hence my bias.

ricardoV94 commented 1 year ago

ya incorporate!

ColtAllen commented 1 year ago

I'm planning to work on https://github.com/pymc-labs/pymc-marketing/issues/134 soon. The methodology for adding time-invariant covariates is rooted in survival analysis:

https://www.brucehardie.com/notes/019/time_invariant_covariates.pdf