For Activation, a Directional LTV Model Beats an Accurate One

Most predicted-LTV (pLTV) projects die because they're graded on the wrong exam. The team chases a high R², the model refuses to cooperate, and everyone concludes "the signal isn't there." But for the most common use case, feeding lookalike and value-based audiences to platforms like Meta, you almost never need an accurate point estimate of a customer's future value. You need a directional ranking good enough to carve the population into a handful of segments whose average value is ordered and separated.

That reframing changes everything downstream: which metric you optimize, which features you're allowed to use, where in the funnel you score, how you turn a continuous score into buckets, and how often you refresh. This article walks the whole chain, including the two traps that catch almost everyone: mechanical leakage from the first purchase, and scoring the wrong population.

Scope. This is about pLTV for activation: seeding lookalike and value-based audiences, where the consumer of your model is a platform's own algorithm. It is deliberately not about every use of predicted LTV. If you are valuing a company, forecasting revenue, planning unit economics, or setting per-customer bids, you need a calibrated estimate, and accuracy is exactly the bar. The narrower, load-bearing claim here: when a platform's audience model consumes your output, ranking beats pricing.

Here's the trap. You frame pLTV as a regression problem (predict a customer's N-month value) and you grade it with R² or MAE. You iterate on features, the number crawls up a few points, and eventually you plateau at something unimpressive. Someone senior asks "is 0.05 R² even usable?" and the project loses air.