回归结果专业总结

Executive summary

• The logistic regression model for registration conversion shows good discrimination (AUC=0.79) and solid explanatory power for a behavioral model (McFadden pseudo R^2=0.28). Calibration is acceptable (Hosmer–Lemeshow p=0.37; Brier=0.16).
• All reported predictors are statistically significant and directionally consistent with product intuition. Key levers with the largest practical impact are CTA variant B, shorter onboarding time, more feature trials, push notifications enabled, referral traffic, and returning users.

Model quality

• Sample size: N=32,000.
• Discrimination: AUC=0.79 indicates the model distinguishes converters from non-converters well.
• Fit: McFadden pseudo R^2=0.28 suggests meaningful explanatory power relative to a null model.
• Calibration: Hosmer–Lemeshow p=0.37 shows no evidence of major miscalibration; Brier=0.16 indicates moderate average prediction error. Recommend reviewing calibration plots by decile.
• Controls: Estimates net of site-by-day effects and device type.

Key drivers (odds ratios and interpretation)

• CTA variant B: OR=1.42 (p=0.002). Switching to CTA B increases conversion odds by 42% versus CTA A.
• Onboarding time (per minute): OR=0.92 (<0.001). Each additional minute reduces odds by 7.7%. Example: +5 minutes → OR=0.67 (−33%); +10 minutes → OR=0.45 (−55%).
• Feature trial count (per additional feature tried): OR=1.23 (<0.001). Trying 3 more features → OR≈1.88 (+88%).
• Push notifications enabled: OR=1.31 (p=0.005). Opt-in is associated with 31% higher odds.
• Traffic source (vs. Organic):
  • Paid: OR=1.15 (p=0.03).
  • Referral: OR=1.36 (p=0.01).
• User type — Returning: OR=1.49 (<0.001). Returning users have 49% higher odds than new users.

Interpretation notes

• Odds ratios are multiplicative on odds, not absolute probabilities. Effects can compound across factors.
• The intercept (−2.10) reflects the reference profile at baseline categories and zero values for continuous variables and is not directly a business baseline rate.

Recommendations

• Product and UX
  • Roll out CTA variant B broadly, confirm via A/B test for causal validation.
  • Streamline onboarding to reduce time-to-completion; prioritize steps with highest time cost.
  • Nudge users to try more features early (e.g., guided tours, checklists); target +2 to +3 features tried in the first session.
  • Optimize the push opt-in prompt timing and messaging to raise enablement rates.
• Growth and channel mix
  • Lean into referral programs; they outperform organic and paid in conversion odds. Maintain paid where ROI is positive, but scrutinize subchannel performance.
  • Tailor flows for returning users to capitalize on their higher propensity, while designing targeted interventions to lift new-user conversion.
• Model deployment and monitoring
  • Use probability thresholds aligned to operational capacity or expected value (cost of outreach vs. value of conversion). Calibrate thresholds via validation curves.
  • Monitor discrimination (AUC/PR), calibration (reliability plots, Brier), and drift by traffic source, device, and time.
  • If needed, apply post-hoc calibration (isotonic or Platt) to fine-tune probability estimates.

Limitations and next steps

• Potential endogeneity: Longer onboarding time may reflect friction; address via process changes or instrumental tests. Consider causal A/B tests for UX changes.
• Interactions and nonlinearity: Explore interactions (e.g., OnboardingTime by TrafficSource; FeatureTrialCount by UserType) and possible nonlinear effects (splines for time).
• Stability: With N=32k and strong signals, estimates are likely stable; still, confirm via k-fold cross-validation and out-of-time validation.
• Reporting: Provide segment-level lift charts and calibration plots; include decision curves to align thresholding with business value.

Bottom line

• The model is reliable for ranking and prioritization. The highest-impact, actionable levers are adopting CTA B, reducing onboarding time, increasing early feature engagement, and improving push opt-in, with additional gains from referral traffic and tailored flows for returning users.

中文摘要（商务简报）

模型概览

• 方法与因变量：分客群 OLS；因变量为周销量（件）。
• 样本与拟合：N=1,200（Segment A=600；Segment B=600）；调整R^2=0.64。
• 整体显著性：F=145.2，p<0.001。
• 诊断：残差近似正态；VIF<3；未见严重异方差，结果具有统计稳健性。

分客群关键结果与解释

• Segment A（价值导向）
  • 价格（元）：+1 → 销量−1.8 件（p<0.001）。
  • 社媒渠道占比（个百分点）：+1 → 销量+0.9 件（p=0.004）。
  • 促销深度（个百分点）：+1 → 销量+0.6 件（p=0.02）。
• Segment B（价格敏感）
  • 价格（元）：+1 → 销量−3.2 件（p<0.001），对价格更敏感。
  • 搜索渠道占比（个百分点）：+1 → 销量+1.1 件（p=0.003）。
  • 促销深度（个百分点）：+1 → 销量+0.4 件（p=0.08；5% 水平不显著，10% 水平边际显著）。

跨客群主效应与交互

• 广告投放（AdSpend，$千）：+1 → 销量+0.5 件（p=0.001）。
• 交互（Price*Promo）：系数+0.3（p=0.02）。促销越深，价格对销量的负向影响被部分缓解。举例：促销深度每增加10 个百分点，价格对销量的边际影响向上调整约+3 件/元。

业务含义

• 价格：B 段价格敏感度高，应谨慎提价；A 段相对缓和。
• 渠道：A 段优先提升社媒渠道占比；B 段加强搜索渠道。
• 促销：对 A 段效果显著，对 B 段证据较弱；但与价格存在正向交互，可用于缓冲价格调整带来的销量压力。
• 广告：对两段均有稳定正效应。

建议（基于统计结果）

• 分段定价测试：A 段先行小幅提价试点；B 段以价格稳定为主，必要时配合更深促销以减弱价格负效应。
• 渠道结构优化：A 段增加社媒投放与渠道份额；B 段提高搜索投放与份额。
• 促销策略：A 段可适度加深促销；B 段聚焦定向促销并与价格联动做 A/B 测试。
• 方法与监控：持续追踪残差与方差稳定性；作为稳健性检查，补充稳健标准误与样本外验证。说明：模型为线性关联，不直接推断因果。

Resumen en español (estilo ejecutivo)

Visión general del modelo

• Método y variable dependiente: OLS segmentado; ventas semanales (unidades).
• Muestra y ajuste: N=1.200 (A=600; B=600); R^2 ajustado=0.64.
• Significancia global: F=145.2, p<0.001.
• Diagnóstico: residuos aproximadamente normales; VIF<3; sin signos de heterocedasticidad severa.

Resultados clave por segmento

• Segmento A (orientado al valor)
  • Precio (RMB): +1 → −1.8 uds (p<0.001).
  • Cuota canal Social (p.p.): +1 → +0.9 uds (p=0.004).
  • Profundidad de promoción (p.p.): +1 → +0.6 uds (p=0.02).
• Segmento B (sensible al precio)
  • Precio (RMB): +1 → −3.2 uds (p<0.001), mayor sensibilidad.
  • Cuota canal Search (p.p.): +1 → +1.1 uds (p=0.003).
  • Profundidad de promoción (p.p.): +1 → +0.4 uds (p=0.08; no significativo al 5%, señal al 10%).

Efectos transversales

• Inversión publicitaria (AdSpend, $ miles): +1 → +0.5 uds (p=0.001).
• Interacción (Precio*Promo): +0.3 (p=0.02). Mayor profundidad de promo atenúa el efecto negativo del precio. Ej.: +10 p.p. en promo ajusta el efecto del precio en ~+3 uds por RMB.

Implicaciones

• Precio: sensibilidad mayor en B; A más moderada.
• Canales: priorizar Social en A; reforzar Search en B.
• Promociones: efectivas en A; evidencia más débil en B; la interacción sugiere utilidad para amortiguar ajustes de precio.
• Publicidad: efecto positivo y consistente en ambos segmentos.

Recomendaciones

• Pruebas de precios por segmento: pilotos de aumentos en A; en B mantener precio y combinar con promociones cuando sea necesario.
• Optimización de mix de canales: más Social en A; más Search en B.
• Estrategia de promociones: aumentar en A; en B focalizar y testear con interacción precio*promo.
• Monitoreo metodológico: añadir errores estándar robustos y validación out-of-sample. Nota: resultados describen asociaciones lineales; no implican causalidad directa.