2026/2027
Наука о данных в маркетинговой аналитике
ID 1174008
Статус:
Маго-лего
Кто читает:
Департамент менеджмента
Где читается:
Санкт-Петербургская школа экономики и менеджмента
Когда читается:
1, 2 модуль
Охват аудитории:
для своего кампуса
Преподаватели:
Ляпин Илья Викторович
Язык:
английский
Кредиты:
6
Контактные часы:
48
Course Syllabus
Abstract
The course trains students to apply data-science methods to marketing and customer analytics on real and synthetic datasets, with a focus on decisions that maximize incremental profit under modern measurement constraints (privacy regulation, deprecation of third-party identifiers, attribution loss). The course is built around three pillars:
• (1) Experimentation and causal inference as the foundation of measurement;
• (2) Predictive customer analytics — customer lifetime value, churn, response, time-to-event;
• (3) Personalization and decision systems — uplift targeting, recommendation, contextual bandits.
The working stack is Python-first (pandas, scikit-learn, statsmodels, econml, causalml, pymc-marketing, lifetimes, lifelines, scikit-survival), with R used selectively where it remains best-in-class (notably Meta Robyn for MMM). All assignments are submitted as reproducible Git repositories. Prerequisites: an introductory course in statistics or econometrics and basic programming experience.
Learning Objectives
- • Design and analyze marketing experiments under realistic operational constraints, including variance reduction (CUPED), stratification, sample-ratio-mismatch detection, and sequential testing. • Build and evaluate causal and predictive models that drive marketing decisions, choosing the appropriate method for the available identification strategy. • Measure the incremental effect of marketing actions across channels, using both attribution methods and incrementality-based approaches (geo-experiments, synthetic control). • Develop predictive models of customer behaviour — lifetime value, churn, response probability, time to next event — and translate them into targeting decisions. • Design and evaluate personalization systems that combine uplift estimation, recommendation, and bandit-style exploration. • Produce reproducible analyses (Git, environment management, code review) that meet industry standards for data-science work.
Expected Learning Outcomes
- • LO1. Plan an A/B test end-to-end: define the metric, compute MDE, choose a stratification and variance-reduction scheme, monitor SRM, and apply sequential or fixed-horizon decision rules
- • LO2. Estimate the incremental contribution of paid-media channels using modern marketing mix modeling (Bayesian MMM with PyMC-Marketing or LightweightMMM). • LO3. Critically compare attribution-based and incrementality-based measurement, and design geo-experiments or synthetic-control studies when randomization is not available.
- • LO5. Build customer-lifetime-value models in both contractual and non-contractual settings (BG/NBD, Pareto/NBD, gamma-gamma), and extend to ML-based approaches. • LO6. Build calibrated classification models for churn and response, choose decision thresholds aligned with business objectives, and evaluate them with lift and uplift curves.
- • LO4. Estimate heterogeneous treatment effects (uplift) using meta-learners (S/T/X/R), causal forests, and DoubleML, and translate uplift estimates into targeting policies.
- • LO7. Build time-to-event models for marketing (Cox PH, Random Survival Forest, DeepSurv) and link survival predictions to retention strategy. • LO8. Implement and evaluate recommender systems and contextual bandits for marketing personalization, and explain when each is appropriate.
Course Contents
- Data manipulation in Python (with a tidyverse bridge in R)
- Experimentation: A/B testing for marketing decisions
- Marketing Mix Modeling (MMM): Bayesian and modern
- Attribution and incrementality measurement
- Uplift modeling and heterogeneous treatment effects
- Customer Lifetime Value modeling
- Churn and response prediction
- Time-to-event models for customer analytics
- Personalization and decision systems: recommendation and bandits
Assessment Elements
- Final ProjectEnd-to-end analysis on a real or realistic marketing dataset, submitted as a reproducible Git repository (code, environment file, README) accompanied by a 4–6 page written business memo. Two project tracks are offered: (i) experiment analysis with variance reduction and heterogeneous-effect estimation, or (ii) uplift-based targeting policy with offline evaluation.
- In-class Assignments
- QuizzesWeekly short quizzes on material from the previous week's lecture.
Interim Assessment
- 2026/2027 2nd module0.35 * In-class Assignments + 0.45 * Final Project + 0.2 * Quizzes
