Promotional Uplift Modelling

Promotional uplift modelling is the practice of estimating the incremental sales a promotion generates — the difference between what actually sold under promotion and what would have sold anyway (the baseline). RELEX defines promotion forecasting as predicting how campaigns, discounts and promotions affect demand by analysing historical sales, market trends and consumer behaviour, with the gap between the promotion forecast and the baseline forecast being the promotional lift/uplift (Web — Promotional Uplift Modelling 2026-06-25). It is the analytical core that feeds both supply-side decisions (how much to buy and position via S&OP (Sales & Operations Planning), Open-to-Buy (OTB) and replenishment) and demand-side decisions (which customers and discount depths actually create new demand rather than giving away margin).

---

Why it matters

• RELEX, in partnership with Incisiv, reports that promotions drove $1 trillion in sales in 2023, that in 2023 promotions drove ~20% of sales in both the U.S. and EMEA, and that 87% of retailers anticipated maintaining or increasing promotional efforts (as-of 2023; per RELEX/Incisiv report published 2025-01-09). (Web — Promotional Uplift Modelling 2026-06-25)
• Despite the scale, RELEX states 37% of retailers surveyed across the US, UK, France and Germany still use spreadsheets as their core platform for managing promotions (as-of 2025-01-09). (Web — Promotional Uplift Modelling 2026-06-25)

Decomposition of a promotion's effect

Tredence frames a full promotion-effectiveness decomposition into baseline, uplift, subsidization, cannibalization, pull-forward, and halo effect, each calculated separately (Web — Promotional Uplift Modelling 2026-06-25).

Key terms

Term	Meaning (per source)
Baseline	Demand that would have occurred without the promotion
Uplift / lift	Incremental sales above baseline attributable to the promotion
Subsidization	Margin given to customers who would have bought anyway
Cannibalization	Sales pulled away from other (often similar) products during the promo
Pull-forward	Demand brought forward in time (pantry-loading / forward-buying), producing a post-promotion dip
Halo effect	Lift on related products driven by the promoted item

Source: Tredence, "Calculating Promotion Effectiveness: A Deep Dive" (2024-06-10). Formulae on the page render as images and were not extractable as text. (Web — Promotional Uplift Modelling 2026-06-25)

Cannibalisation & halo: how they are modelled

• RELEX states cannibalisation and halo relationships are identified from receipt-level transactions and loyalty data via association rule learning, or from SKU-store-level time series by analysing correlations when promotions break the normal "equilibrium" (negative correlation = cannibalisation, positive = halo). The halo effect can be modelled as a weaker special-case promotion and cannibalisation as a promotion with negative effect. (Web — Promotional Uplift Modelling 2026-06-25)
• RELEX also notes that correctly attributing cannibalisation/halo to promotions improves the baseline (non-promo) forecast too, because the ML model stops mistaking those biases for baseline demand. (Web — Promotional Uplift Modelling 2026-06-25)

Causal / ML methods

• DoorDash presented "Causal Machine Learning for Promotions: Industry Evidence and Applications" at KDD 2025 (Toronto, Aug 2025), describing a two-stage framework: (1) estimate each customer's true incremental response, then (2) optimize which promo to deliver under budget/eligibility constraints. (Web — Promotional Uplift Modelling 2026-06-25)
• DoorDash names the core problem the "non-incremental promotion" problem: many discounted orders would have happened anyway, so blanket discounts erode margin by paying "always buyers" instead of "persuadables" — "the promotion didn't create new demand — it simply gave away margin unnecessarily." It used Double Machine Learning (DML) for both discrete treatments (which promo to give) and continuous treatments (how deep a discount). (Web — Promotional Uplift Modelling 2026-06-25)
• In DoorDash's reported results, causal-targeted campaigns achieved nearly the same incremental orders as a uniform baseline at roughly half the cost per incremental order; a continuous discount-depth model beat a uniform discount on both order-rate lift and cost efficiency in a large-scale A/B test (as-of 2025-12-16; vendor/practitioner self-report). (Web — Promotional Uplift Modelling 2026-06-25)

[!unverified] A KDD 2025 causal-ML-for-promotions paper reportedly shows Double Machine Learning achieving the strongest out-of-sample results (Qini score 8.03%, Uplift@30 1.11%). This is from a search snippet only — the PDF was not fetched/verified. Standard uplift toolkit (Uplift Trees / Random Forests, S-/T-/X-Learner meta-learners; Uber's open-source CausalML) is similarly snippet- level. (Web — Promotional Uplift Modelling 2026-06-25)

How it feeds demand planning, S&OP and replenishment

RELEX describes the integration chain: an ML promo uplift forecast (per store/channel, factoring promo type, discount depth, display, seasonality, weather, competitor activity) → automated store replenishment with configurable early-delivery proportions → DC forecasts built from store-order projections → virtual ringfencing to protect online inventory; with forward-buy ("investment buy") optimisation traded off against inventory carrying cost, shelf life and timing (Web — Promotional Uplift Modelling 2026-06-25). In the monthly S&OP cycle, promotional uplift is confirmed during the demand review step (see S&OP (Sales & Operations Planning)).

Vendor consolidation reflects this convergence: per Lokad's competitor commentary, ToolsGroup acquired Evo (late 2023) to align pricing/promotions with inventory, and o9 markets Revenue Growth Management integrating RGM, demand planning, supply chain and IBP on one platform (Web — Promotional Uplift Modelling 2026-06-25). [Lokad authors competitor reviews — flag bias.]

Benchmarks (as-of 2025-01-09)

• RELEX claims modern supply chain solutions can improve promotional forecast accuracy by 15% (relative; vendor self-report). (Web — Promotional Uplift Modelling 2026-06-25)
• RELEX cites German drugstore chain Rossmann achieving a 30% reduction in inventory for promoted items while improving availability, plus a 10% reduction in stock-outs on promoted items (vendor case study; case date not stated). (Web — Promotional Uplift Modelling 2026-06-25)

Fashion / apparel & DTC dynamics

[!unverified] DTC-specific signals were single-sourced from marketing blogs and not deep-fetched — treat as directional only: reportedly 76% of DTC companies planned deeper discounts in 2025 and 69% raised marketing spend; apparel/fashion gross margins ~50–65%; DTC apparel CAC reportedly rose 222% over 8 years (24.7% inflation in 2025); deep discounting is said to "train" high-value buyers to wait for markdowns (the discount-addiction / cannibalisation risk that connects to Markdown Optimisation). (Web — Promotional Uplift Modelling 2026-06-25)

The tension here links to the S&OP (Sales & Operations Planning) finding that in DTC, demand is partly controlled via paid-media spend — making promotional and media levers an explicit demand input rather than something planners only respond to.

Contradictions / tensions

What practitioners report

[!unverified] No Reddit practitioner signal was gathered this run — the Reddit MCP was unavailable (see Reddit — Promotional Uplift Modelling 2026-06-25). YouTube provided video-level pointers only (no transcripts; see YouTube — Promotional Uplift Modelling 2026-06-25). Both are recorded as gaps to be re-harvested.

Open frontiers from this page

• Cannibalisation & Halo Effects — modelling mechanics; quantified magnitudes still thin
• Causal Inference for Pricing & Promotions — DML, meta-learners, uplift trees
• Post-Promotion Dip — pull-forward / pantry-loading magnitude and duration
• Retail Media Incrementality — adjacent: 71% of advertisers reportedly now rank incremental lift as their most important retail-media KPI (as-of 2025; single source, not deep-fetched)
• Trade Promotion Management (TPM) — manufacturer/wholesale promo planning angle