Forecasting promotions: baseline, uplift and cannibalisation
A promotion forecast separates what would have sold anyway from what the promotion adds and what it takes from other items. Each step needs its own data and its own caution.
KEY TAKEAWAYS Summary by the editors
- A promotion forecast has three parts: the baseline (sales without the promotion), the uplift (extra sales caused by it) and the cannibalisation (sales taken from substitute products).
- Cannibalisation is the share of a promoted item's volume uplift that is taken from substitute products, and it is usually measured in units rather than revenue because volume is closer to consumer choice.
- In a published simulation study, a better baseline and a longer time window both reduced cannibalisation estimation error, and smoothing-based baselines were roughly 40 to 50% better than a naive last-value baseline.
- Any new forecasting method should be compared with simple benchmarks such as the naive and seasonal naive forecasts, and judged on data it has not seen.
- Fashion adds difficulty because styles are short-lived, so there is often little clean history for the same item without a promotion.
Forecasting a promotion means estimating three things separately: the baseline, which is what would have sold without the promotion, the uplift, which is the extra volume the promotion causes, and the cannibalisation, which is the volume it takes from other products. Most forecasting errors in promotions come from mixing these three together, so a good method keeps them apart.
This explainer describes each component, the data needed, how to measure accuracy and where machine learning helps and does not help. It uses published academic and reference sources and avoids claims about specific retailer results.
What is a promotion baseline?
The baseline is an estimate of sales in the promotion period if no promotion had run. It is never observed directly, so it has to be modelled. Uplift is then the difference between actual sales and that baseline. If the baseline is wrong, every uplift number built on it is wrong too.
A simple baseline takes the last value before the promotion. A study of cannibalisation from Aalto University used exactly this naive approach, taking the last pre-promotion value and comparing it with the first promotion week, and noted that better options such as linear interpolation with smoothing or Prophet would perform better. In its simulation, smoothing-based interpolation was roughly 40 to 50% better than the naive baseline. The study used simulated data, so the size of the gain will differ with real data, but the direction is useful: baseline quality matters a great deal.
For fashion, a baseline is harder than in grocery. Styles are short-lived, sizes sell unevenly, and a promoted item may never have had a non-promoted period. Teams then borrow information from similar items, earlier seasons or the early weeks of the same style.
How is promotional uplift estimated?
Uplift is estimated by comparing promoted sales with the baseline, usually in units. It depends on discount depth, the promotion mechanism (percentage off, bundle, free shipping), channel, timing, marketing support and stock availability. A model learns these effects from past promotions, so it needs a record of each promotion with its dates, mechanics and scope.
- Record every promotion with start and end dates, depth, mechanism, channel and affected items.
- Record stock availability, because a stockout in a promotion week looks like weak uplift when it is actually lost sales.
- Record overlapping activity, such as campaigns, email sends and wholesale-driven discounts at retail partners.
- Keep a clean non-promoted history for the same or comparable items, since this is what the baseline learns from.
What is cannibalisation, and why does it matter?
Cannibalisation is a drop in a product's sales caused by the company's own activity. Reference definitions describe it as lost sales volume, revenue or share caused by the same company launching a new product, and they list discounting as one example: lower prices on one product pull buyers away from pricier alternatives, and the effect fades when prices return to normal.
The Aalto study defines it more precisely for promotions: the share of a promoted item's volume uplift taken from substitute products. It works in volume rather than turnover, because volume is more directly tied to consumer choice. Negative values indicate complementarity, meaning two items sell together. For a merchandiser, this means a promotion that shows a strong uplift on one style may be moving sales between styles rather than creating new demand.
| Component | Question it answers | Main data | Main risk |
|---|---|---|---|
| Baseline | What would have sold without the promotion? | Non-promoted history, seasonality, stock | Too little clean history for short-lived styles |
| Uplift | How much extra volume does the promotion cause? | Promotion records, depth, mechanism, channel | Stockouts and overlapping campaigns distort the signal |
| Cannibalisation | What does it take from other items? | Item-level sales across substitutes, assortment structure | Error grows with more products and more noise |
How do machine learning methods estimate cannibalisation?
The Aalto paper describes an elastic net regularised alternating least squares method that estimates a cannibalisation matrix between item pairs. It masks uplifts for non-promoted items, constrains self-cannibalisation to zero and uses a validation check to limit overfitting. In simulations the method converged to the same solution regardless of the starting guess, but error grew with more products and higher noise. The authors report a runtime of about two hours for 40 products over three years of real store data on a laptop, and they note that real retail data will probably give less accurate results because consumer behaviour departs from the model's assumptions.
The sober reading is that item-to-item cannibalisation can be estimated, but not precisely, and not for very large assortments without simplification. Many teams therefore estimate it at the level of product groups, or use it only to identify the strongest cannibalisers rather than to compute exact numbers, which is the use the authors themselves suggest.
How should a promotion forecast be evaluated?
Accuracy should be judged on data the model has not seen. The forecasting textbook by Hyndman and Athanasopoulos recommends a test set of about 20% of the sample, at least as long as the longest forecast horizon, and warns that a good fit to training data does not guarantee good forecasts. It also lists the weaknesses of percentage errors: MAPE is undefined when an actual value is zero, unstable near zero and penalises negative errors more heavily than positive ones. Scaled errors such as MASE divide by the error of a naive forecast, so a value below 1 beats the naive benchmark.
The same book states that any new forecasting method should be compared against the naive and seasonal naive methods, and that if it cannot beat them it is not worth considering. For promotions, a useful extra benchmark is the planner's own estimate, which leads to the forecast value added approach described in our how-to on that topic.
What are the limits for fashion?
- Short life cycles leave little history for the same item, so models borrow from similar items and earlier seasons.
- A position paper on learning-based forecasting in fashion notes that classical methods such as ARIMA cannot be fitted when there is no history or only a few observations, which is common for new styles.
- Wholesale-led brands often cannot see the promotions their partners run, so the sell-out signal is incomplete.
- Promotions interact with each other and with clearance activity, so effects measured in one period may not carry to the next.
Used with these limits in mind, promotion forecasting is a way to make planning assumptions explicit: a stated baseline, a stated uplift and a stated cannibalisation estimate that can be reviewed and corrected after the event.
Frequently asked questions
What is the baseline in promotion forecasting?
It is the estimate of sales during the promotion period if no promotion had run. It cannot be observed, so it is modelled from non-promoted history, seasonality and stock. Uplift is measured against it, so errors in the baseline pass straight into the uplift figure.
What is cannibalisation in retail promotions?
It is the share of a promoted item's volume uplift that is taken from substitute products. It is usually measured in units rather than revenue. A strong uplift on one item can therefore hide a drop in similar items.
How accurate are machine learning promotion forecasts?
It depends on data quality and assortment size. A published simulation study found that error grew with more products and more noise, and that real data will likely give less accurate results than simulations. Any model should be tested against naive benchmarks on unseen data.
Which accuracy measure is best for promotions?
No single measure is best. MAPE has known weaknesses with zeros and small values, while scaled errors such as MASE compare the model with a naive forecast. Many teams report a scaled error and a bias measure together.
One edition every weekday morning. Read in five minutes. Free for industry professionals.
SOURCES
- Aalto University: Estimating cannibalization between item pairs from sales data
- Wikipedia: Cannibalization (marketing)
- Hyndman and Athanasopoulos: Forecasting: Principles and Practice, Evaluating forecast accuracy
- Hyndman and Athanasopoulos: Forecasting: Principles and Practice, Some simple forecasting methods
- arXiv: On the use of learning-based forecasting methods for ameliorating fashion business processes
