M2-011Medium优先级与取舍 · Prioritisation & trade-offs🔥 大厂高频出现 6 次
如何评估一个促销/补贴活动的“增量”效果?
How do you measure the incremental effect of a promotion or subsidy?
💡 考场高分一句话答案:
必须构造反事实:用未受补贴人群作对照(geo/holdout),或用倾向得分匹配,否则会把“本来就会买的人”算成增量。
You must construct a counterfactual: an untreated control (geo or holdout) or propensity matching, otherwise you count people who would have bought anyway.
📌 核心考察要点 (Key Points)
- 核心问题:补贴会带来三类人——增量新客、提前购买(pull-forward)、本来就会买的人(deadweight)
- 设计:随机 holdout(最干净)、geo 对照、或按阈值做 RDD(如满 100 减 20)
- 度量:增量收入 = 处理组收入 − 对照组收入;同时看活动后 30–60 天的“回落”
- 陷阱:pull-forward 会让短期增量虚高,长期净增量为零甚至为负
- The core problem: a promotion reaches three groups — genuinely incremental new buyers, pull-forward buyers, and deadweight who would have bought regardless
- Design: randomised holdout (cleanest), geo control, or RDD around a threshold such as spend 100 get 20 off
- Measurement: incremental revenue = treated revenue − control revenue, plus a 30–60 day post-period to see the payback
- Trap: pull-forward inflates the short-term lift and can make the long-run net increment zero or negative
📐 第一性原理与数理推导机理
净增量应扣除提前购买:$$\text{Incremental}=\sum_{t\in \text{promo}} (Y_t^{T}-Y_t^{C})+\sum_{t\in \text{post}}(Y_t^{T}-Y_t^{C})$$若只看活动期,会系统性高估。活动后对照组反超(因为处理组已提前消费)是 pull-forward 的典型证据。
⚖️ 业务权衡与取舍 (Trade-offs & Calibration)
补贴评估还要看人群异质性:新客补贴的增量通常高,老客补贴的 deadweight 极高。因此最优策略是“对新客/流失客补贴,对活跃老客不补贴”,这也是为什么真实系统里用 uplift model 而不是统一发券。
⚠️ 常见高频面试陷阱与避坑指南
- 用活动期前后对比代替对照组
- 只看活动期,不看活动后的回落
🔍 大厂面试官连环追问 (Follow-ups)
- 如何用 uplift model 定位“补贴敏感人群”?
- 如果增量收入为正但净利为负,怎么办?
🎯 关联大厂考题实战训练 (6 题)
嫌字多看太累?进入苏格拉底式对话推演实验室
零打字负担,以大厂考官真实困境切入,通过点选直觉选项暴露思维盲区,手把手推导 Why。
⚡
Test muscle memory on this topic via 3-Layer Flashcards
Active Recall: Mentally formulate Prompt ➔ Math Derivation ➔ Benchmark Implementation in 3 minutes.