发表机构
Universidad de Huelva（韦尔瓦大学）
机构由 AI 辅助整理，请以论文原文为准。AI 中文总结
本文提出结合双重机器学习和保形预测的因果框架，通过弃权（不执行）诊断处理零售价格弹性估计中的混淆，并在数据不足时分类原因，显著降低估计误差。
AI 中文摘要
本文提出一个因果决策框架，用于估计零售渠道中的价格弹性，该过程通常受到促销、竞争对手行动和市场摩擦的干扰。当数据模糊时，系统不强行计算，而是引入决策弃权（不执行）（\textsc{wait}）作为主动诊断工具，而非估计失败。结合双重机器学习和保形预测，该工具评估是否存在可靠条件来调整价格，或者暂停决策是否更可取。当系统弃权（不执行）时，它会详尽分类暂停原因，识别哪些产品需要设计定价实验，或者将数据聚合到品牌层面是否能恢复可用的估计。在受控合成数据上的测试表明，这种操作纪律大幅降低了估计误差（将RMSE从0.571降至0.159），并为薄数据零售环境中的盲目估计提供了一种实用且安全的替代方案。
英文摘要
This paper presents a causal decision framework for list-price elasticities in intermediated retail channels, where estimates are confounded by promotions, competitor moves and shopkeeper pass-through. Rather than forcing a number when the evidence is ambiguous, the system treats abstention (Wait) as an active diagnostic output. Built on Double Machine Learning and conformal prediction, it evaluates whether conditions exist to act on a recommended price or whether withholding is preferable, and it states when: acting beats withholding exactly when the calibration slope of the true headroom on the recommended move exceeds one half, a condition that predicts the sign of the gain in 37 of 38 synthetic groups. When the system abstains, it classifies the cause and flags which products are candidates for a designed pricing experiment. Aggregating estimates to the brand or category level, where independent price designs average out, cuts the root mean squared error from 0.488 to 0.159, although aggregation alone does not make the intervals honest. Tested on synthetic data with known truth and on public scanner data, the study reports its costs: the system abstains on 52% to 87% of presentations that were identifiable, and a single-window estimate forecasts real price changes poorly (calibration slope 0.22). What it offers is a way to audit any abstention rule and to publish at the unit the evidence supports; no commercial panel or pilot was used, so it does not measure commercial uplift.
