Bayesian Statistics and Causal Inference for SKU‑Level Pricing in E‑commerce

Business background: The e‑commerce platform Zhuangzhuang sells millions of mobile and 3C items daily, requiring automated, market‑responsive pricing to reduce manual effort and improve revenue and user experience.

Ideal pricing model: Beyond global accuracy, the model must avoid extreme mispricing, provide deterministic control, and allow human‑in‑the‑loop adjustments, thus requiring interpretability and controllability.

Previous attempts: (1) Manual rule‑based pricing on limited attributes; (2) Deep regression models using historical sales data (black‑box, prone to bad cases); (3) Hybrid manual‑plus‑rule approaches that still lack fine granularity.

Bayesian and causal inference pricing scheme – fundamentals: Introduces Bayesian inference, prior, likelihood, posterior, and estimation methods such as Maximum Likelihood Estimation (MLE), Maximum A Posteriori (MAP), and Markov Chain Monte Carlo (MCMC) sampling.

Causal inference basics: Defines treatment variable (e.g., coupon issuance), response variable (e.g., transaction), treatment effect (uplift), and covariates. Discusses randomized experiments, propensity scoring, and a deep‑network approach (Dragonnet) that does not require explicit randomization.

SKU‑level pricing with Bayesian statistics: Treats each SKU (same model, capacity, condition) as a unit, models price‑vs‑transaction probability with a logistic distribution, uses MCMC to sample posterior parameters, and derives a controllable price function that can be adjusted based on inventory and sales speed targets.

Product‑level pricing with causal inference: Extends the SKU model by sampling multiple price points as treatment values, training a Dragonnet three‑head network to predict both propensity scores and outcomes, and selecting the price that maximizes the estimated treatment effect while respecting business constraints.

Summary: The proposed framework integrates Bayesian MCMC pricing at the SKU level with Dragonnet‑based causal inference for finer product‑level adjustments, offering both statistical rigor and practical controllability for large‑scale e‑commerce pricing.

References: Includes links to resources on MLE/MAP, PyMC3, MCMC in Python, causal forests, and related research papers.

Author: Fan Hanqin, algorithm engineer at Zhuangzhuang, responsible for pricing strategy development. Contact: fanhanqin