Can Your Demand Forecast Tell You Which Deals to Run?

Accurate on Volume, Split on What Each Deal Added

This series follows one team through a year of pricing and promotion decisions. It starts with trade promotion at a set list price. Later posts take on list price, the store brand, competitors and new packs. I made up Larkfield Foods, Kestrel Market and the people in this story, and the numbers come from a simulated market. Earlier posts covered what happens to sales after a deal ends and how a team's own deals create the soft months it then reacts to. They also covered why post-event reports make each deal look better than it was and how many stores built the displays a deal paid for.

---

When Your Retailer Wants Deals Locked Earlier

Your demand forecast includes a lift for every planned deal. When a retailer wants your deals locked earlier, you may want to use those lifts to choose which deals to run.

Grace Lin, the cereal buyer at Kestrel Market, told Larkfield Foods she would accept no more late adds. Kestrel would lock every promotion slot well ahead, and she wanted Larkfield's deals for next quarter.

Riley Shaw, who heads revenue growth, went to Hannah Weiss, who runs demand planning. "Grace wants next quarter's deals now," Riley said. "Can I choose them from your forecast?"

"It can tell you how many boxes a plan will sell," Hannah said. "First, let me show you how accurate it's been."

What Your Demand Planner Does

Demand planning reports to supply chain. Hannah's team forecasts how many boxes each product will sell at each retailer. The scan data reports sales by week, so the forecast does too. Supply uses that forecast to plan production and shipments. The forecast adds each planned deal's lift to a baseline, what the product sells without a deal. Sales and supply then adjust it each month in sales and operations planning (S&OP), and the key account manager often sets a deal's lift by hand.

Building a forecast takes four steps:

1. Gather the scan sales and the trade plan, past and ahead.

2. Fit a few models that read the season and the plan.

3. Test each model on past sales. Make each forecast with only the sales you had when its slot locked, then compare it with what sold. Planners call this test a backtest and the forecast made at the lock a lag forecast.

4. Keep the best model, and review its forecast with sales before the plan locks.

Most planning tools run steps 2 and 3 for you and call the winner "best fit."

The Models Behind Your Demand Forecast

Beyond a simple seasonal baseline, forecasting models come in three families. Classic time-series models learn a product's pattern from its own history. Machine learning models learn how several inputs together move sales. Pretrained models are the newest family. They train on a large collection of other time series, real and computer-generated, before they read yours. Hannah ran five models on the same sales and the same trade plan:

Model	Family	How it forecasts
Seasonal baseline	Rule of thumb	Recent full-price sales times the usual rise and fall across the year; it ignores deals
SARIMAX	Classic time series	Sales follow from recent sales, plus the effect of the trade plan and the season (the ARIMA family)
Prophet	Classic time series	Sales split into a trend, a yearly season and the effect of the trade plan (Meta's open-source library)
Chronos-2	Pretrained	Reads your sales and plan and forecasts without any fitting to your product (Amazon's open model)
Gradient-boosted trees (GBM)	Machine learning	Dozens of small decision trees, each correcting the ones before it (LightGBM, XGBoost)

Your Forecast Is Close Enough to Plan Supply

Hannah forecast all of last year and the start of the new one, using only the sales the team had when each slot locked. She fixed each model's settings before the check began. She also gave each model the deal that ran in each slot, as if the team had locked it at that point. Last year, the team added most deals later, so a forecast made at the lock couldn't have known them. This check shows how accurate a forecast can be when the team knows each deal at the lock, as it now will.

The gold line is the forecast made when each slot locked, with its deal known (Simulated market).

Demand planners grade a forecast with two standard measures. Forecast error, usually WAPE (weighted absolute percentage error), adds up the misses in both directions and divides by what sold. A lower number is better, and many teams report accuracy instead: 100% minus WAPE. Bias adds up the misses with their signs and divides by what sold. A bias above 0% means the forecast ran high, and a bias below 0% means it ran low. Hannah reported error separately for sales at full price and sales while a deal ran, because a forecast can do well on one and poorly on the other.

Model	Forecast error, at full price	Forecast error, while a deal ran	Bias, overall
Seasonal baseline	15%	64%	−34%
SARIMAX	14%	13%	−3%
Prophet	20%	18%	−3%
Chronos-2	13%	16%	+3%
Gradient-boosted trees	19%	19%	0%

Simulated market.

The seasonal baseline misses the deals by design, because it ignores them. It supplies normal sales, the first part of a planner's forecast: what the product sells at full price, including the dips from your past deals. At full price, its error was no worse than Prophet's or the trees'. Among the four models that read the plan, SARIMAX and Chronos-2 had the smallest errors overall (13.5% and 14.7%). That gap is too small to separate them. The chart shows SARIMAX. Over the whole period, all four totals came within 3% of what sold, although a single quarter could miss by more (the workbook shows SARIMAX's quarters). On a plan that copies last year's deals into next quarter, SARIMAX, Chronos-2 and the trees forecast about 20,500 to 20,800 boxes, and Prophet forecast about 23,200.

What Your Forecast Says Last Year's Deals Added

Hannah selected the 13 deals from last year with at least two months left in the year. For each deal, she went back to when its slot locked. There she asked each model to forecast the deal and the two months after, once with the deal and once without. The difference is what the model says the deal added.

While the deals ran, three of the four models that read the plan estimated the 13 deals added 19,500 to 21,300 boxes, near the post-event reports' 21,700. Chronos-2 estimated 23,900. The models split on the two months after. Once Hannah counted those months too, the estimates ran from 12,600 boxes with Prophet to 18,600 with the trees.

"Each model gives me a different answer for what last year's deals added, thousands of boxes apart," Hannah told Riley. "That range is too wide to choose deals from."

"Why so wide, when three of them agree on what last year's plan would sell next quarter?" Riley asked.

"Most of them are close on what sold while the deals ran, in total. They split on what came after, the dip we keep debating," Hannah said. "Our history doesn't settle it. We booked 28 of our 40 deals because sales looked soft, and we've never gone two months at full price between deals. None of these models has seen what two quiet months after a deal would sell. Our baseline hasn't either, because we set it from full-price sales that follow our own deals."

Hannah kept SARIMAX's forecast for next quarter in the S&OP plan and planned to update it as slots locked. She gave Riley's team the seasonal baseline as their normal sales, the starting number for any deal plan. Riley still had no way to choose the deals Grace wanted.

What Last Year's Deals Actually Added

Your own data can't show what a deal added, because no month runs twice, once with the deal and once without. A simulated market can. Because I built this one, I replayed last year, removing each deal in turn and leaving everything else unchanged. The replay uses expected sales and keeps the same luck with and without each deal. I calibrated the market to published industry figures where they exist, and its model card lists the sources and assumptions.

Simulated market.

While they ran, the 13 deals added about 20,900 boxes, close to the reports and to three of the four models. Over the two months after, shoppers bought about 9,000 fewer boxes than they otherwise would have. More than half of that loss fell while the next deals ran. Shoppers had boxes at home and remembered the lower price. A deal soon after the last one also lifted less. Summed deal by deal, the 13 deals added about 11,900 boxes of Toasted O's 12 oz. On this history, every model estimated more.

Prophet came closest on the total for the two months after, while the trees expected almost no loss. On average, a deal added about 920 boxes. Deal by deal, SARIMAX and Prophet missed by about 320 and 360 boxes, the trees by about 590 and Chronos-2 by about 790. SARIMAX and Chronos-2 forecast sales about equally well, as did Prophet and the trees, yet each pair differed widely on what a deal added. On this product's history, the forecast error table couldn't tell those models apart.

Check Your Own Forecast With This Workbook

Download the forecast-check workbook (Excel) below.

forecast-check.xlsx

Download XLSX • 11KB

1. Open the workbook in Excel, because a phone or email preview may show zeros. It starts with Larkfield's sales from this post, so you can see what each sheet shows.

2. On the Weeks sheet, replace Larkfield's rows with one product at one retailer, one row per week as your scan data reports it, numbered from 1. For each row, enter what sold, the forecast as it stood when the slot locked, whether a deal ran, and why you booked it. If you don't know why, mark the deals that followed a soft patch.

3. Read the Summary. It shows forecast error and bias at full price and on deals, and how many of your deals you booked because sales looked soft.

If you booked most of them because sales looked soft, treat your forecast's deal lifts as untested.

What These Tools Couldn't Tell You

You lock next quarter's deals before any of them sells. Your forecast can tell supply how many boxes to make for a given plan, as Hannah's did with last year's deals copied in. It can't tell you how many boxes a deal truly adds. It learned from sales your own deals shaped, and the models split on the answer.

Next, Riley's team builds the year's calendar on a capped trade budget. It has one other set of lifts to work from: the post-event reports.

How does your team check the deal lifts inside its forecast? Tell me in the comments.

Words Behind Your Forecast

Term	What it means
Demand planning	The supply chain team that forecasts what each product will sell, so plants build to it
Sales and operations planning (S&OP)	The monthly meeting where sales, marketing, finance and supply agree on one plan
Demand forecast	The boxes a product will sell, from past sales and the plan
Lag forecast	The forecast your plan held at a set time before the sales it covers, kept to check accuracy
Backtest	Checking a forecast on past sales, made from only the data you had at the time
Forecast error (WAPE)	Weighted absolute percentage error: the misses either way, added up and divided by what sold
Bias	The misses with their signs, added up and divided by what sold; above zero means the forecast ran high
Late add	A deal added to a retailer's calendar after the plan locked
Post-event report	A read of a deal's sales against a baseline of normal sales, often the sales just before it; the difference while the deal runs is the deal's lift
Soft patch	A stretch when sales looked low against recent sales, the reason the team booked many deals (the workbook's "Soft month")