You run a 25% sitewide sale starting Friday. You ordered inventory based on your baseline forecast. By Saturday afternoon, three of your top five SKUs are sold out. Sound familiar? The problem isn't the sale. The problem is that you planned for a normal week.
Why Baseline Forecasts Fail During Promotions
A baseline forecast answers one question: what will sell if nothing unusual happens? It accounts for seasonality, day-of-week patterns, and trend. What it doesn't account for is the fact that a 20% discount on a $60 item can shift demand by 2 to 4x overnight.
Most operators know this intuitively. But knowing it and actually modeling it are two different things. When purchase orders go out three to six weeks before a sale, the person placing those orders is usually looking at the same demand signal they always use. The promotion is on the calendar, but the forecast doesn't reflect it. The result: under-ordered SKUs that stockout in hour 36, and a lot of "we should have bought more of that" conversations after the fact.
Promotion lift forecasting closes that gap. Instead of layering a rough "add 30%" multiplier on top of baseline, it uses historical event data to estimate expected uplift per SKU, per promotion type, before the purchase order is placed.
How Lift Modeling Actually Works
The core mechanic is straightforward: look at what happened the last time you ran a similar promotion on a similar product, and use that as your lift estimate.
In practice, this means maintaining an event log. Every promotion gets tagged with key attributes: discount depth, promotion type (sitewide vs. category vs. SKU-specific), channel mix (email blast, paid ads, organic), duration, and calendar context (was it around a holiday? a competitor sale?). When a new promotion is planned, the model finds historical events that are reasonably similar and computes the average demand multiplier observed during those windows.
That multiplier gets applied to the baseline forecast for the sale period. If your baseline says 80 units on a normal Friday, and historical data shows a 3.2x lift for 20%-off sitewide sales in spring, the lift-adjusted forecast becomes 256 units. That's the number your replenishment logic should be working from, not 80.
In our experience, the biggest forecasting errors happen when operators apply a single blanket multiplier across all SKUs. A bestseller behaves very differently from a slow-moving item during a promotion. Bestsellers often see a smaller relative lift because demand is already near its ceiling. Slow movers can spike dramatically with any promotional attention. The model needs to operate at the SKU level, not the category or site level.
Categorizing Promotions for Better Estimation
Not all promotions lift demand the same way. Applying a single "promotion multiplier" is better than nothing, but it leaves a lot of accuracy on the table. Here's a practical way to think about promotion categories for lift estimation purposes:
| Promotion Type | Typical Lift Range | Key Variable |
|---|---|---|
| Sitewide discount (15-25%) | 1.8x to 3.0x | Email list size, ad spend |
| Flash sale (24-48 hrs, deep discount) | 3.0x to 5.5x | Urgency messaging, channel saturation |
| Category sale (single vertical) | 2.0x to 3.5x | Category size, SKU depth in category |
| Bundle or BOGO | 1.5x to 2.5x (volume) | Bundle pair demand elasticity |
| Loyalty/member-only sale | 1.2x to 2.0x | Active loyalty base size |
These ranges are starting points, not targets. Your actual multipliers will be tighter or wider depending on your customer base and how aggressively you promote. The goal is to build your own table from real event data over time, not borrow benchmarks from someone else's industry report.
The New SKU Problem: No Promotion History
Here's where things get genuinely hard: what do you do with a SKU that has never been promoted before? Or worse, a new product in a category you've never discounted?
You have no promotion history. The model has nothing to anchor on.
A few approaches that work reasonably well in practice:
- Proxy SKU method: Find the most structurally similar product in your catalog that has been promoted. Same price range, same category, comparable unit velocity. Use its historical lift as the prior estimate for the new SKU, with a conservative haircut of 15-20%.
- Category baseline method: Use the average lift observed across all SKUs in that category during previous promotions. If you've never promoted that category, step up one level to department or sitewide averages.
- Channel-weighted prior: If the promotion is primarily email-driven and you know your email conversion lift historically, use that as a floor estimate. It won't be SKU-specific, but it gives you a directional signal.
None of these are perfect. That's the honest reality. For genuinely new SKUs with no analogues, conservative over-ordering is usually the right call. Stockout cost, in terms of lost revenue and customer frustration, typically exceeds the carrying cost of holding a few extra weeks of safety stock.
A good rule of thumb: for new SKUs in a first-time promoted category, target 1.5x your baseline forecast as a starting inventory position, then update your model once you have one real event in the books.
What Happens When You Skip the Lift Adjustment
The math is unforgiving. A 40% stockout rate during a promotional window doesn't just mean 40% of potential revenue disappeared. It means disappointed customers who were already primed to buy. Some will backorder. Most won't. Some will find the product elsewhere and not come back.
We've tracked operators who run sitewide sales without lift-adjusted forecasts, and the pattern is consistent: they sell through 60-70% of their promotional inventory in the first 18 hours, stockout on key items by hour 30-36, and end the sale period with a mixed result. Strong top-line numbers, but with 15-25% more revenue left on the table than necessary.
That's not a demand problem. It's a planning problem. The customers showed up. The inventory didn't.
Building the Habit: Event Logging as Operational Infrastructure
The most important thing you can do right now, before you try to build a formal lift model, is start logging every promotional event you run. Date range, discount structure, promotion type, channel, and actual units sold vs. baseline for each SKU. That's it. Even a simple spreadsheet works.
After 6 to 8 events, you'll have enough data to start making lift estimates that are meaningfully better than guessing. After 15 to 20 events, you'll have a real model. The compounding value of that event log is hard to overstate. Every promotion you run without logging it is data you'll never get back.
Fact: operators who maintain structured event logs for 12 months typically reduce promotional stockouts by 30-40% compared to their first year of operating without them.
The operational shift here isn't complicated. It's not about sophisticated machine learning or building a data engineering team. It's about treating promotions as structured experiments with observable outcomes, and capturing those outcomes systematically so the next decision is better than the last one.
Key Takeaways
- Baseline forecasts don't account for promotion lift. Using them for pre-sale ordering is a structural under-order waiting to happen.
- Lift estimates work by finding historically similar promotional events and computing the average demand multiplier observed.
- Categorize promotions by type, depth, and channel. A sitewide 20% sale and a 48-hour flash sale are not the same signal.
- For new SKUs with no promotion history, use proxy SKUs or category averages with a conservative buffer.
- Start logging events now. The model improves with each data point. Waiting for a complete dataset means you never start.
Planning before a sale isn't a luxury for large brands. It's the operational baseline for anyone running promotions with real inventory exposure. The sale won't wait for your supply chain. Your forecast needs to move first.
