英文标题：A Synthetic Benchmark Dataset with Endogenous Marketing Spend for Validating Marketing Mix Models
作者：Niklas Heusch
arXiv ID：2608.21130 | 分类：stat.AP | 发表：2026-09-07
许可：CC-BY
摘要 营销组合模型（MMM）从观测时间序列中估计广告的增量销售效应，然而它们很少针对真实基准进行验证，因为在真实数据中真实基准是不可观测的。合成数据原则上可以弥合这一差距，但现有的生成器以{{NL}}外生方式生成营销支出——忽略了估计问题的核心{{NL}}难点，因为真实预算是围绕促销日历、季节性和近期表现来规划的。本{{NL}}文提出了一个参数化生成器，以及一个固定的参考{{NL}}实例，用于一个合成的每周零售数据集（156周，三个媒体{{NL}}渠道），其中支出源自四种有据可查的协调{{NL}}机制——季度预算反馈、在{{NL}}促销日历之前的预期性支出、计划性的电视爆发投放，以及算法化的绩
Marketing Mix Models (MMMs) estimate the incremental sales effect of advertising from observational time series, yet they are rarely validated against ground truth, because ground truth is unobservable in real data. Synthetic data closes that gap in principle, but existing generators produce marketing spend exogenously - omitting the central difficulty of the estimation problem, since real budgets are planned around promotional calendars, seasons, and recent performance. This paper presents a parameterized generator, and a fixed reference instance, of a synthetic weekly retail dataset (156 weeks, three media channels) in which spend arises from four documented coordination mechanisms - quarterly budget feedback, anticipatory spending ahead of a promotional calendar, scheduled TV bursts, and algorithmic performance chasing - on a demand baseline with seasonal, quality, price, and unobserved sentiment components. Spend translates into incremental sales through two transformations, carryover and diminishing returns, instantiated here as geometric adstock and logistic saturation with known parameters; the true causal decomposition of every week's sales is recorded alongside the error-contaminated variables a practitioner would observe. Every mechanism is a parameter that can be varied or switched off, and a companion procedure simulates go-dark geo-experiments with exact treatment effects. The seeded generator and reference instance are publicly released with notebooks that reproduce every number in this paper.
查看完整双语翻译 →
正在跳转到翻译阅读页… 如果没有自动跳转，请点击这里。