MMM 101 - Rapid ROI Onboarding Course
Unilever MMM Program · 00. Learning Material
All tracks
Course homeWelcomeBuilt 2026-07-12 from the Learning Material source pack · CURRICULUM.md v2026-07-11
MMM 101: the Rapid ROI onboarding course
This course takes anyone joining the Unilever MMM program - fresh analyst through experienced hire - from "what is marketing mix modelling" to "I can own a cell". It is authored fresh from this account's own source material: the Rapid ROI Pre-Read primer, the MathCo methodology workbook, five KT session transcripts, the production model config, and the master file's own vocabulary. Modules cite their sources at the foot.
Pick your track:
- Fresh analyst - read everything in order: Track 0, then 1, 2, 3, then the capstone.
- Experienced MMM hire - start at Track 2 (this account's solution). Dip back into Track 1 only where a gap shows up; module 2.3 is your bridge if your background is Bayesian MMM rather than frequentist state-space.
- Program Lead / PM - Track 0 alone is a complete 60-minute read. Add 2.2 (cell lifecycle) and 2.5 (deliverable pipeline) if you want the two most PM-relevant deep modules.
Modules marked Interactive carry live charts (drag the sliders - the data is synthetic and regenerates identically on every load). Modules marked SME review pending were authored ahead of the SME interviews and carry explicit check-markers wherever content needs confirming; the full list is in the review register below.
Course scope
23 modules
4 tracks + capstone · interactive charts · check-yourself quizzes · program-exact vocabulary
Phase 1 of the onboarding plan
House rules baked into every module
- No real client numbers anywhere - every exercise uses synthetic data.
- Program vocabulary matches the master file exactly (stages, statuses, feed types).
- Real production artifacts are pointed at as exhibits, never excerpted.
Modulesin reading order per track
Track 0MMM in 60 Minutes
MODULE 0.1
What MMM Is (and Isn't)
MMM vs MTA vs incrementality, and the four client questions Rapid ROI answers
MODULE 0.2
Adstock and Saturation, Intuitively
Why media effect lags and decays, and why doubling spend rarely doubles sales
Interactive
MODULE 0.3
What This Program Delivers
The deliverable arc, the KPI family, and the cell lifecycle vocabulary
Track 1Foundations
MODULE 1.1
CPG and Media Taxonomy
Product hierarchy, media hierarchy, ATL vs BTL, sell-in vs sell-out
MODULE 1.2
Data Requirements per Feed Type
The 14 feeds, the readiness lifecycle, and the sign-off gates
MODULE 1.3
From Regression to MMM
The core equation, why plain OLS breaks, additive MMX vs multiplicative PnP
MODULE 1.4
EDA and the Data Summary Deliverable
Why the EDA gate exists and how to write a data summary in program style
MODULE 1.5
Transformations: Adstock and Saturation
The math behind the intuition, and how the production YAML encodes it
Interactive
MODULE 1.6
Bound Setting I: Base Variables
Defensible LB/UB ranges for price, distribution, promo, and macro
MODULE 1.7
Bound Setting II: Media ROI, ASSR, Halo, Competitor
ASSR as the benchmark, LTCF, halo rules, competitor scaling
MODULE 1.8
KPIs and Decomposition
ROI, mROI, CPM, Effectiveness, ASSR, Contribution % vs Due-to %
Track 2This Solution
MODULE 2.1
The Rapid ROI Model Architecture
ST and LT hierarchies L1-L4, online/offline splits, fallback rules
MODULE 2.2
Cell Lifecycle End to End
Three phases, ten stages, the 8-week Build clock, and cell_status semantics
MODULE 2.3
Model Core: UCM, State-Space, and Kalman Filtering
Why this program's engine is a structural time-series model, not plain OLS
InteractiveDeep dive
MODULE 2.4
Long-Term Modeling: MDS and LTCF
Meaningful / Differentiation / Salience, and converting ST ROI to ST+LT
MODULE 2.5
The Deliverable Pipeline: Workbook to Deck
Tracing one deliverable from analytical workbook to client story
SME review pending
MODULE 2.6
Config and Platform Setup in Practice
Assembling a real mmx_tool_config YAML block by block
Track 3Advanced
MODULE 3.1
Model Validation and QA
Statistical pass/fail criteria, business-KPI checks, and the model scorecard
SME review pending
MODULE 3.2
Error Allocation
Distributing a total-level residual down to platform level
MODULE 3.3
Optimization and Scenario Simulation
Response curves, portfolio and flighting optimization, the mROI rule in action
InteractiveSME review pending
MODULE 3.4
Common Failure Modes and Troubleshooting
The recurring pitfalls this program's own trainers flag
MODULE 3.5
Industry Context: Bayesian MMM vs This Program
Meridian, Robyn, and where our UCM/frequentist approach sits
Deep dive
Capstone
SME Review Registerauto-compiled from the check-markers inside modules
Open items to confirm with the named SMEs
The three planned interviews (Gopi/Omkar on optimization, Pijush on validation, insight storytelling) had not happened when this course was built - these markers are exactly what to ask them.
| Module | SME | What to check |
|---|---|---|
| Rajesh / Tushar | The Pre-Read's Effectiveness label says "per 10 Impressions" but its formula multiplies by 10 against a CPM defined per 1000 - confirm the intended units convention (revenue per 10 impressions vs per mille) before teaching it as gospel. | |
| Rajesh Kumar Yerra | The mid-pack price default of -10% to -12% comes from a garbled transcript passage ("10 to 12"), interpreted as a starting magnitude for a mid-position brand - confirm the intended band before teaching it as a fixed rule. | |
| Rajesh Kumar Yerra | The theory sheet's high-correlation distribution band (35-45%) sits above the KT session's typical leader set-point (30-35%, observed 20-45) - confirm which anchors current practice, and whether the promised correlation-by-market-position dictionary supersedes both. | |
| Rajesh Kumar Yerra | Applying the LT-share/ST-share definition to the e-com 85/15 split gives an LTCF well below paid social's stated 0.55-0.6, which inverts the expected funnel ordering - confirm the intended per-channel LTCF table (the fuller version is said to live in the bound-setting template) before teaching exact values. | |
| Rajesh Kumar Yerra | The methodology sheet states different bands from the KT sessions: competitor variables at 60-85% of the own-base analogue (vs KT's 1/4 to 2/3 with a 2/3 hard cap) and halo at 5-15% of the analogous media variable (vs KT's 5-10% max ~10%) - confirm which bands are current practice and whether the sheet predates the KT refinements. | |
| Program Lead | The Process sheet's 28-step team-handoff SOP predates current tbl_Stages naming - confirm whether it should be reconciled to the current stage vocabulary for teaching or retired as historical context only. | |
| Shirsha | Confirm that Harvey's structural time-series (UCM) formulation is the actual academic lineage of the production engine, so the citation can be taught as "where our method comes from" rather than "the nearest textbook". | |
| Rajesh Kumar Yerra / Harsha MN | The bands above are the values stated aloud in the KT(3)/KT(4) sessions; the trainers referenced a fuller per-channel, per-market-type LTCF table living in the bound-setting template - confirm the complete table (and the remaining channels' splits) before this module is treated as the reference. | |
| Program Lead / storytelling SME | This entire insight-layer section is authored from general consulting practice, not from program source material - the dedicated insight-storytelling SME session has not yet happened (SOURCE_MAP gap 5), so confirm these principles match how this program actually writes and reviews client decks, and replace or extend them with the program's own rules once that session lands. | |
| Shirsha | The curriculum's module 1.5 assessment premise "a narrower beta range implies stronger priors on that platform's saturation" does not match the production config (beta is one identical grid for all platforms; only alpha differentiates) - confirm the reality taught here and reword that assessment item. | |
| Pijush | Confirm how the score_df composite ranking actually weights or sequences these criteria in practice - whether priority_order is a strict lexicographic sort, a weighted composite, or a filter-then-rank scheme is not documented anywhere internal. | |
| Pijush | The DW comfort band, the BP p-value cutoff, and the VIF ceiling actually enforced in practice are not written down anywhere internal - the table below uses textbook conventions (DW roughly 1.5-2.5, BP p > 0.05, VIF < 10) which need confirming or correcting against what the team really applies. | |
| Pijush | Confirm the walked pass-vs-fail example planned for the SME session - a real (anonymised) candidate pair where one passes statistically but fails a business lens would anchor this module far better than the synthetic table below. | |
| Pijush | The Methodology workbook's Check List sheet is a 2-row stub (only "Data: sales data present" and "ST L1: variable signs correct" are populated) - confirm whether a real, complete working QA checklist exists elsewhere that should replace it as this module's reference artifact. | |
| Rajesh / Tushar | Confirm which formula is operative in production: the Error Allocation Theory sheet's text says factor = Predicted/Actual while its own worked example computes parent-contribution / sum-of-children (1.1236 above) - the two coincide only in the special case where the children's sum equals the parent's predicted value, so the discrepancy needs an authoritative resolution. | |
| Gopi / Omkar | Confirm whether optimization runs on ST response curves only or on ST+LT combined curves - the RC_Curve_Data exports exist on both sides of the model, and whether long-term (brand-health-routed) returns enter the allocation objective changes recommendations materially for upper-funnel channels. | |
| Gopi / Omkar | Confirm what solver the production optimizer actually uses (grid/greedy over the RC step tables, scipy-style constrained optimization, or something else) - the equalize-marginal-ROI account above is the standard method, not a description of the code. | |
| Gopi / Omkar | Confirm which constraint set is standard for this account (floors/ceilings? max percentage change vs last year? per-BG or per-platform constraints?) - the list above is generic MMM practice, not this program's documented default. | |
| Gopi / Omkar | Confirm how flighting optimization is actually implemented in this program (a weekly-grain optimizer over the fitted transforms? heuristic guidance off the alpha/beta parameters? a separate tool?) - nothing internal documents it beyond the session title. | |
| Gopi / Omkar | Confirm how simulation scenarios are packaged for clients on this account (a standing simulator tool, a deck section with fixed scenario sets, ad hoc runs on request?) and how many scenarios a standard read-out carries. | |
| Program Lead | The Phase 2 notebook set and the synthetic dataset generator are still to be built - when they land, this capstone page should be re-pointed at the actual notebook paths and the step numbering below reconciled against what was implemented. |
Who to askfrom Training Plan_MMX.xlsx (April bootcamp roster)
| Topic area | SME |
|---|---|
| Marketing measurement fundamentals, Media Mix vs Marketing Mix | Ashutosh |
| Data sources and model design | Sujit, Arka |
| MMX KPIs (base, price/promo, media metrics) | Rajesh, Tushar |
| Media model construct (frequentist, additive/multiplicative, constraints, priors) | Shirsha |
| Price and Promotion (PnP) model construct | Diya |
| Model validation (topline, due-to story, ROI story, CPM, mROI) | Pijush |
| Standard deliverables (media principles, adstock, reach penetration, landscape) | Tanima, Mausami |
| Halo, synergy, cannibalisation, reach and response curves | Haishma, Bhuvaneshwari |
| Optimization and simulation | Gopi, Omkar |