Marketing and Growth Analytics Questions

ML approach to channel-level ROI and LTV prediction

Features

• Touch-level: channel, touch_time, touch_position (first/last/mid), cost_at_touch.
• Customer-level aggregates: counts per channel, time-weighted exposures, recency, first-touch channel, source mix.
• Downstream: time-to-opportunity, conversion outcome, ACV, retention metrics.

Model choice

• Two-stage approach: (1) uplift/attribution model (e.g., additive hazard or multi-touch fractional attribution via Shapley) to allocate conversions across channels; (2) customer LTV model (e.g., survival + GBM or probabilistic LTV like Pareto/NBD + spend regression) using allocated credit per channel as features.

Validation & lookalike bias

• Use causal inference: propensity-score stratification or doubly-robust estimators controlling for pre-treatment covariates to reduce selection bias.
• Holdout experiments if available; otherwise, use temporal or geo-based holdouts. Compare modeled channel ROI to experiment-derived lift when possible.
• Stability checks: reweight channels, perform sensitivity to unobserved confounding (E-value), and check that predicted high-LTV channels align with experiment signals.