Machine Learning for Growth Marketing

Welcome: the week ahead and how we work

Genial Labs · The Day 1 opening

Five days, five questions

  1. Day 1. Who are our customers, and who is still with us?
  2. Day 2. What is a customer worth, in money, with uncertainty?
  3. Day 3. Did the marketing cause the sales?
  4. Day 4. Where should the next dollar go?
  5. Day 5. Whom should we target, and what is the plan for next quarter?

Module 0 (Pre-work: environment, warm-up and coding agents) is optional pre-work before Day 1.

Every module ends in a decision

Machine learning for growth marketing is only useful if it changes what someone does on Monday.

Whom to retain · what a customer is worth · what to spend where · whom to send the discount

  • Each lab closes with its decision cell, made from your own results.
  • The same retailer and its data come back all week, so each method is measured against the one before it on the same customers.

The path

  1. Pre-work: environment, warm-up and coding agents

  2. Transaction logs to customer-base analysis

  3. Buy-till-you-die models: BG/NBD and Pareto/NBD

  4. Covariates in CLVTools, and BTYD versus machine learning

  5. Monetary value and dollar CLV

  6. From CLV to decisions

  7. Experiments done right

  1. Geo-experiments and quasi-experiments

  2. Bayesian marketing mix models

  3. Calibrating and validating an MMM

  4. Budget allocation under uncertainty

  5. Uplift and heterogeneous treatment effects

  6. Capstone: one decision pipeline for the retailer

Each step fixes a failure of the one before it.

How the parts connect

Three lanes ending in decisions: geo tests to measured lift to a Bayesian MMM to the allocator; transactions to BTYD and Gamma-Gamma to CLV; experiments to uplift models to targeting. CLV feeds the allocator's long-run objective and the targeting margin.

  • Customer value (middle): transactions → BTYD and Gamma-Gamma → CLV.
  • Media and lift (left): geo tests → measured lift → priors or likelihood for the MMM → the allocator.
  • Targeting (right): experiments → uplift models → who gets the offer.
  • CLV connects them: the allocator’s long-run objective and the targeting margin.

How a module runs

BriefingLab in ColabDebrief

  • Briefing: why the method is needed, and how it works, at the depth you need to use and check it.
  • Lab: you build it in a notebook. Debrief: the room compares results and states the decision.
  1. Predict what a cell will show. Write it down.
  2. Run it.
  3. Explain the result, especially a surprise.
  4. Check it against a checkpoint, often against the known truth.

Working in the labs

  • Each exercise is a function stub with its solution folded beneath it. Try first; open the solution when you are stuck.
  • A checkpoint after each exercise passes or fails and says whose code it checked: yours or the reference.
  • Stuck on exercise N: run its solution cell, then workshop.use_reference(N) (Python) or use_reference(N) (R), and carry on.
  • Run all stops at the first exercise you have not written. WORKED_EXAMPLE runs the whole lab on the reference solutions.
  • The stretch section at the end is optional. Nothing later depends on it.

The honesty rules

  • The readiness page says what has run. Where, when, with which settings. A run time appears only if a run record exists, and a laptop or CI run is never shown as a Colab run.
  • Ground truth first. Synthetic data come from seeded generators with known parameters. Checkpoints compare your estimate with the truth using a named interval, such as a 94% highest density interval.
  • No invented facts. Every reference was fetched and is dated. No made-up statistics, people or quotes. When we do not know, we say so.

Setup: nothing to install

  • Every lab runs in Google Colab, in the browser. You need a Google account.
  • Most labs are Python; some modules add an R notebook where R has the reference implementation.
  • The install cell installs the package versions the labs were tested with. If it asks you to restart the session, restart, then Run all again.
  • Prefer your own machine? The Setup page covers the workshop’s Docker image.

Logistics

  • Each day starts at 09:00 with a warm-up and ends with a wrap-up; the breaks and lunch are on the schedule.
  • Days 2 and 3 have an afternoon clinic on that day’s decision.
  • Day 5 ends with the capstone: pairs build one decision pipeline and present a short brief.

Where everything is

Next

Module 1 · Transaction logs to customer-base analysis

Who are our customers, and who is still with us?