Generalized Causal Forest: Construction and Application in Online Trading Markets

Online trading has become ubiquitous, prompting the application of machine‑learning techniques to achieve supply‑demand balance; a key challenge is accurately estimating the price‑demand curve, which is often confounded by factors such as seasonality, heterogeneity, and high‑dimensional data.

The presentation outlines four parts: background introduction, existing algorithms, construction of the generalized causal forest (GCF), and experiments with deployment.

Background : The rapid growth of online transactions creates a need for models that capture the nonlinear, potentially negative relationship between price and demand while controlling for confounders. Difficulties include nonlinear curve shapes, heterogeneous effects across locations, seasons, and user groups, and the massive scale of data.

Existing Algorithms : Industry‑popular methods assume a piecewise‑linear price‑demand relationship, which is often unrealistic. Causal forests extend random forests for causal inference by employing three key mechanisms: (1) sample splitting (e.g., bootstrap) to obtain unbiased leaf estimates, (2) honest estimation using separate data for tree structure and outcome estimation, and (3) splits that maximize conditional treatment effect heterogeneity (CATE/CAPE).

Generalized Causal Forest Construction : GCF estimates the dose‑response function nonlinearly using kernel regression (e.g., Gaussian kernel) to weight nearby observations, defines a PDRF distance based on first‑order derivatives to quantify heterogeneity, and incorporates robust methods that require only one of propensity‑score or outcome models to be accurate. The SPARKGCF algorithm implements these ideas for big‑data environments.

Experiments and Deployment : Simulation results show GCF outperforming several baseline models. In a ride‑hailing pricing system, GCF is deployed such that when supply exceeds demand, discount strategies are applied to stimulate demand, improving order completion rates. The deployment architecture and reference materials are also presented.

Overall, the generalized causal forest provides a flexible, unbiased, and scalable solution for estimating heterogeneous treatment effects in large‑scale online trading scenarios.