Applying Causal Inference to Financial User Operations: Scenarios, Challenges, and Practices

Introduction This article introduces the use of causal inference in financial user operations, aiming to help practitioners quickly apply causal models to their business scenarios.

Typical Scenarios

1. Programmatic advertising – rising ad costs motivate the integration of causal models to make ad placement more precise and efficient, potentially delivering significant economic benefits.

2. User operations – in telemarketing, the phenomenon of “lying orders” (users assigned to a sales rep before contact) can be mitigated by causal models that identify natural conversions and allocate resources more effectively.

3. Additional scenarios such as brand advertising, effect advertising, activation, retention, revenue, and referral growth are discussed, emphasizing the relevance of “what‑if” questions across the AARRR framework.

Challenges

Data challenges include the scarcity of random samples, overlapping interventions, long conversion cycles, and missing key features that are essential for causal modeling.

Business challenges arise from new touchpoints (push, WeChat, AI voice) that lack historical data and from evolving gamified or socialized operation strategies.

Causal Sample Library – built using Propensity Score Matching (PSM) and business‑system data to support robust causal modeling.

PSM‑PS stage : compute propensity scores for historical data and perform matching; the propensity score represents the probability P(T=1|X=x) of an individual receiving the treatment.

PSM‑Matching stage : select appropriate sample pairs, set thresholds, and evaluate data distribution and business metrics.

Sample library construction steps include scoring, matching, intelligent‑marketing data selection, intervention mapping, feature assembly, weighting, efficient storage, monitoring of treatment/control distributions, and training causal models.

Business Practice

Common questions address the existence of “sleeping dogs” (negative response groups), offline metric thresholds for model deployment, and experiment design.

Experiment scheme : random experiments with control and test groups to measure delta business effects, uplift limits, and identify sleeping dogs.

Confounders are factors causing P(Y|X) ≠ P(Y|do(X)). Mitigation strategies include merging business‑system features, modeling other strategies, and adding uplift model features.

Technical Overview shows the end‑to‑end pipeline: after building the causal sample library, budget‑aware sample selection, intervention mapping, and feature assembly feed into optimization components and causal models, which are then deployed via the intelligent marketing platform to conduct uplift experiments.

Full‑Link Empowerment aims to reduce acquisition cost in advertising, improve second‑day retention, increase credit approvals, boost ROI in user operations, and enhance referral growth.

Q&A

Q1: Can uplift be done without a control group? A: Use an alternative treatment (e.g., SMS) as a proxy control.

Q2: Does the causal sample library depend on specific marketing risk targets? A: The library is built generically; targets are linked later through data assembly.

Q3: How to map treatment in the same dimension? A: Example – map coupon usage to an adjusted interest rate.

Q4: Should extreme propensity scores be trimmed? A: Yes, removing highest and lowest scores can reduce anomalies.

Q5: Minimum size of the PSM sample library? A: Business‑level metrics require millions of records; the full library often needs an order of magnitude more.

End of the presentation.