Growth ML: Causal Incentive Allocation

Core claim

The problem is causal allocation, not targeting.

The strongest explanation starts with the decision being made, then works backward to the data and modeling structure needed to support that decision.

Core idea

We estimate the incremental value and incremental cost of each eligible action, evaluate candidate policies with logged randomized data, and allocate incentives under budget and efficiency constraints.

What makes this hard

The same design has to satisfy causal validity, ranking quality, budget constraints, and online measurement. Optimizing one part while ignoring the others produces brittle gains.

1Identify

Use randomized assignment and propensity logs to define causal labels.

2Learn

Estimate heterogeneous value and cost with calibrated uplift models.

3Decide

Prune invalid actions, rank by marginal economics, and optimize.

4Validate

Use IPS, doubly robust OPE, and online tests before scaling.

5Repeat

Keep exploration alive so future policies remain identifiable.

Concept map

How the pieces connect.

This is the structure behind the allocation problem. The model is only one part of the policy.

Each piece exists to prevent a specific failure. Randomization prevents confounding, separate heads prevent objective leakage, OPE prevents offline overclaiming, pacing prevents budget mismatch, and monitoring catches drift.

Starting point

Why heuristic campaigns break down.

A rule-based incentive can be easy to operate but still waste spend because it lacks counterfactual measurement and user-action personalization.

Heuristic policy

Eligibility is based on coarse lifecycle rules.
Offer depth is mostly fixed or manually tuned.
Reporting focuses on observed redemption or conversion.
Budgeting is separated from user-level marginal value.

Causal allocation policy

Eligibility is paired with randomized holdouts.
Offer choice is action-specific and user-specific.
Measurement targets incremental value and incremental cost.
The decision layer optimizes under explicit constraints.

The reframing

This is not a conversion prediction problem. It is an economic treatment-assignment problem: which action creates incremental value, for which user, at what incremental cost, under which budget?

Formal problem

Define the estimand before the model.

The policy is valuable only if the estimated quantities match the decision being made.

          Valueτv(x,a) = E[Yv(a) - Yv(0) | X = x]
          Costτc(x,a) = E[Yc(a) - Yc(0) | X = x]
          EfficiencyCPIV(x,a) = τv(x,a) / τc(x,a)
        

Outcome value

Use the value definition that matches the business decision. A training proxy can be valid if it preserves allocation quality and is reconciled to the accounting metric.

Incremental cost

Cost is not the nominal offer value. It is the cost caused by assignment, which can differ from expected liability, deposited credit, or realized spend.

No-offer action

No-offer must be represented as an action because forced treatment leaks budget into users whose best policy is to receive nothing.

          ChoosemaxSx,a ∈ {0,1} ∑x,a Sx,a τ̂v(x,a)
          Subject∑x,a Sx,a τ̂c(x,a) ≤ Bperiod
          ConstraintCPÎ(x,a) = τ̂c(x,a) / τ̂v(x,a) ≤ threshold
        

Identification

Randomization gives the causal comparison.

If historical assignment reflects intent, observed outcomes combine user state, treatment choice, and business rules. Randomized assignment makes causal labels possible.

Binary treatment design

For a fixed campaign variant, randomize eligible users between treatment and holdout. This identifies campaign-specific heterogeneous effects for value and cost.

τ(x) = E[Y(1) - Y(0) | X = x]

Multi-action design

For a menu of actions, log the probability of each assigned action. Offline policy evaluation needs the full action propensity, not just a treatment flag.

a ∈ A ∪ {no offer}

Common failure modes

Action-space explosion: too many actions thin out support and raise treatment-effect variance.
Long-horizon labels: outcome windows that are too long can become too noisy for ranking.
Interference: concurrent incentive programs can contaminate holdout interpretation.

Modeling

Separate value and cost because they are different statistical problems.

A good design treats model architecture, feature scope, labels, and calibration as product choices, not notebook choices.

The optimizer consumes estimates. If value and cost have different noise, density, or calibration behavior, one shared score can hide the failure mode.

Why uplift learners

They target heterogeneous treatment effects directly and can handle treatment-control imbalance better than naive conversion models.

Feature strategy

Segment where behavior and identification are coherent. Lifecycle-specific models can outperform a global model early in a program's life.

Selection criteria

Prioritize incremental calibration, monotonic ranking in holdout bins, and split stability over raw predictive fit alone.

Labels

Separate the learning horizon from the accounting horizon.

Longer outcome windows can be better for business reporting and worse for model training if they dilute the rank signal.

Learning horizon

The horizon used to train and rank candidate actions. It should match the intervention's response dynamics and preserve incremental ordering.

Accounting horizon

The horizon used for planning, reporting, and financial reconciliation. It can be longer if it is not forced into the high-variance training target.

Practical choice

When value labels are sparse, I would rather train on a stable proxy that ranks interventions well and then reconcile the selected policy to the longer-run business metric through calibration and validation.

Decision layer

Predictions become assignments only after pruning, ranking, optimization, and pacing.

The decision layer is where economics and budget constraints become an allocation policy.

1Score

Estimate incremental value and cost for eligible user-action pairs.

2Prune

Remove negative-value or poor-efficiency pairs before optimization.

3Rank

Order surviving pairs by marginal economics, not raw conversion.

4Allocate

Maximize value under budget, efficiency, and eligibility constraints.

5Pace

Adjust available budget using actual spend and expected liability.

Pruneτ̂^v(x,a) < 0 or CPÎ(x,a) above threshold PacingB_available = B_period - actual spend - expected remaining liability

Best phrasing

The business chooses the budget and efficiency constraints. The model ranks candidate actions so the policy can allocate budget to the highest-value cases.

Offline policy evaluation

Evaluate the policy, not just the model.

A model metric may look good while the policy it induces performs poorly under the logged action distribution. OPE makes the policy comparison explicit.

IPSV̂_IPS(π) = (1/n) ∑_i [1{W_i = π(X_i)} / p̂(W_i|X_i)] Y_i DRV̂_DR(π) = (1/n) ∑_i [μ̂(π(X_i),X_i) + w_i(Y_i - μ̂(W_i,X_i))]

IPS dependency

Requires known propensities and sufficient overlap. It is direct, but can have high variance when propensities are small.

DR advantage

Combines an outcome model with propensity correction. It can reduce variance and gives another lens on policy value.

Online gate

OPE is a screening step, not a substitute for an online experiment when interference and implementation effects matter.

Exploration

Data generation is part of the product.

An adaptive policy needs continued exploration. If exploration is skewed toward promising areas, logging and correction become mandatory.

Warm-start exploration is not a new estimand. It is a sampling policy plus correction through logged propensities.

Risk

Skewed exploration can make naive averages biased and can starve low-probability actions of support.

Mitigation

Log propensities, enforce minimum exploration where needed, use IPW or DR estimators, and watch overlap diagnostics.

Portfolio effects

Interference is structure, not noise.

When multiple incentive programs touch overlapping audiences, the measured effect of one program can depend on what the other programs are doing.

What goes wrong

Holdout interpretation becomes conditional on the portfolio state.
Per-program lift can be biased if other assignments are invisible.
Policies can compete for the same marginal user behavior.

What to build

Log every assignment that can affect the outcome.
Add competing-treatment features when possible.
Partition audiences when stacking control is unavailable.
Move toward coordinated experimentation across programs.

Evaluation

A useful policy has to pass several checks.

The right question is not "does the model fit?" It is "does the induced policy create incremental value under the constraints it will face?"

1Label calibration

Cost and value separately.

2Ranking separation

Monotonic bins and split stability.

3Optimization checks

Budget, thresholds, edge cases.

4Policy value

IPS and DR versus baseline policy.

5Online validation

Experiment results and monitoring.

Lessons

Design rules worth remembering.

These are the high-yield lessons that generalize across growth, ads, marketplace, and lifecycle ML problems.

#	Lesson	Mechanism	Design rule
1	Long horizons are hard to identify.	Incremental signal gets sparse and noisy.	Train on stable ranking proxies; report longer-run outcomes separately.
2	Large action menus fail without support.	Sample per action shrinks and effect estimates destabilize.	Start compact, then expand around promising regions.
3	Cost and value need separate diagnostics.	They have different label density and variance.	Separate heads and calibration workflows.
4	Campaign averages hide heterogeneity.	Some user-action edges can be efficient even when the average is not.	Optimize at the user-action level.
5	No-offer is a first-class action.	Forced treatment leaks budget.	Include no-offer in the policy action space.
6	Exploration policy affects the estimator.	Skewed assignment changes sample geometry.	Log propensities and correct with IPS or DR.
7	Static thresholds decay.	Offer mix, behavior, and costs drift.	Recalibrate and monitor continuously.
8	Feature parity matters.	Offline-online mismatch degrades policy quality.	Enforce feature contracts and serving monitors.

Equation sheet

The core notation.

Keep the direction of every metric explicit. The important question is whether the quantity matches the decision.

          Effectτ(x) = E[Y(1) - Y(0) | X = x]
          Valueτv(x,a) = E[Yv(a) - Yv(0) | X = x]
          Costτc(x,a) = E[Yc(a) - Yc(0) | X = x]
          CPICPI(x,a) = τc(x,a) / τv(x,a)
          CPIVCPIV(x,a) = τv(x,a) / τc(x,a)
          Policya*(x) = arg maxa scorea(x)
          Scorescorea(x) = valuê(x,a) - λ cost̂(x,a)
        

Common questions

Questions the design has to answer.

These questions separate causal allocation from plain conversion prediction.

How is this different from propensity-to-buy targeting?

Propensity predicts what a user is likely to do. Uplift estimates what the action changes. For incentives, that distinction matters because high-propensity users may convert without treatment, while lower-propensity users may have high incremental response.

The decision should optimize incremental value net of incremental cost, not raw conversion probability.

Why do you need randomized holdouts?

Historical assignment is usually endogenous. The business chose who to treat based on lifecycle state, spend intent, or risk. Without randomization, observed outcomes mix the user state with the effect of treatment.

Holdouts create the counterfactual labels needed to estimate incremental value and cost.

Why not maximize predicted value directly?

Because incentives consume scarce budget. A high-value action can still be inefficient if it is too expensive. The policy needs constrained optimization: maximize incremental value subject to budget and marginal efficiency thresholds.

Why can offline model metrics be misleading?

The metric may not match the policy objective. A model can have decent prediction error but induce bad allocations if it misorders marginal users or is miscalibrated in high-spend regions.

Offline policy evaluation uses logged propensities to estimate the value of the actual policy being considered.

What would you monitor after launch?

I would monitor assignment volume, spend pacing, realized cost, value calibration, rank monotonicity, feature drift, action propensities, no-offer rate, overlap diagnostics, and online experiment results. The point is to monitor both the model and the allocation policy.

Final review

Review checklist.

Use this as a compact summary of the technical argument.

State the estimand first.

Incremental value, incremental cost, and policy value before model details.

Explain no-offer.

No-offer prevents forced spend and gives the optimizer a real outside option.

Defend the logging schema.

Actions, propensities, outcomes, and competing assignments are required for causal evaluation.

Separate model metrics from policy metrics.

Calibration and ranking matter, but IPS and DR evaluate the policy.

Avoid private specifics.

Use symbolic budgets, generalized examples, and conceptual diagrams in public settings.