What Is Demand Forecasting?
Demand forecasting is the estimate of how many units you will sell, by SKU and by channel, over a stated period. The four methods CPG brands actually use, what data each one needs, the conditions that break them, and how to tell whether yours is working.
·
9 min readEvery buying decision you make rests on one number you cannot observe: what you are going to sell. Demand forecasting is how you produce that number deliberately instead of by feel, and how you find out afterwards whether it was any good.
What is demand forecasting?
Demand forecasting is the process of estimating how many units of each product customers will buy over a future period. The inputs are sales history, the events you already know about, and the seasonality the history reveals. The output is units — by SKU, by channel, by week or month.
The level matters more than most guides admit, because a forecast is only useful at the level of the decision it feeds:
| The question you are answering | The level you need | The horizon |
|---|---|---|
| What do I order from this supplier now? | SKU × channel, weekly | Longer than your lead time |
| Can we fund the Q4 buy? | Units × unit cost, monthly | Two to three quarters |
| Do we need a second co-packer? | Category, monthly | A year or more |
| Should we say yes to this retail deal? | SKU × retailer × door, weekly | Launch, plus the replenishment run |
A total-company revenue forecast answers none of those. Neither does a SKU-week forecast that stops before your lead time does. Getting the level wrong is the most common failure here, and it looks like accuracy right up until someone tries to place an order against it.
What does a demand forecast actually decide?
The forecast is upstream of nearly every operational number you use:
- Safety stock is sized from how much demand and lead time vary around the forecast, not from a flat number of weeks of cover. See what safety stock is.
- Reorder points are the forecast for the length of your lead time, plus that buffer. See what a reorder point is.
- Purchase order quantities are the forecast over your order cycle, then pushed up or down by supplier constraints like minimum order quantities.
- Cash planning is the forecast multiplied by unit cost and dated by when each order has to be placed, which is usually well before the revenue arrives.
- Production and co-packer capacity is the forecast rolled up to the level your co-packer schedules at.
- Retail commitments are the forecast for a launch that has not happened yet, which is the hardest version of the problem.
Get the forecast wrong low and you stock out, lose the sale, and teach next year's model that demand was lower than it was. Get it wrong high and the cash sits on a pallet. Both failures are expensive, but only the second one is visible on a balance sheet, which is why most brands discover they have been under-forecasting long after it started.
Which forecasting method should you use?
There are four methods that cover almost everything a consumable CPG brand needs. The useful question is not which is best but which fits each SKU:
| Method | What it needs | Where it works | How it fails |
|---|---|---|---|
| Moving average | A few months of history | Steady, mature products with no strong season | Lags every turn. It cannot see a peak coming, only that one has arrived |
| Exponential smoothing (ETS) | A few months of history | Products with a gradual trend, up or down | Overreacts to a one-off spike and carries it forward |
| Seasonal models (SARIMA and similar) | About a year of history, ideally more | Anything with a repeated annual pattern — holiday, summer, back-to-school | Invents a season if you feed it less than a year, and mistakes last year's stockout for last year's demand |
| Regression | A measurable driver, such as price or promotion depth | Products whose demand is driven by something you control and record | Falls apart when the driver is not recorded, or when the relationship changes |
Two practical points follow from that table. First, a catalogue almost never wants one method — the hero SKU with three years of history and the launch SKU with six weeks of it are different problems. Second, the model choice matters less than the data you feed it; a well-chosen model on uncleaned history will lose to a moving average on clean history every time.
The mechanics of each method, with the arithmetic, are in statistical forecasting methods for CPG. If you are choosing a period rather than a model, weekly versus monthly forecasting is the more useful decision to make first.
How do you build a forecast, step by step?
1. Pull the history at the level you will act on
Pull units, not revenue, at SKU × channel × week. If you plan to place separate orders for Amazon and DTC, forecast them separately and add them up — a blended number hides the channel that is actually moving. Multi-channel demand planning covers where the channels genuinely differ.
2. Clean the history before you model it
The single highest-value step, and the one most often skipped. Exclude or correct:
- Stockout weeks. Sales were capped by availability, not by demand. Left in, they teach the model that nobody wanted the product during the weeks nobody could buy it.
- One-off spikes. A viral post or a one-time bulk order is not a repeating pattern.
- Channel changes. The week you turned on a marketplace is a structural break, not growth.
3. Pick a method per SKU
Sort the catalogue by the shape of its history: steady, trending, seasonal, or driven by something you control. Assign accordingly, and expect the assignment to change as products mature.
4. Layer in what the model cannot see
A statistical model only knows what already happened. It does not know about the promotion you have booked, the retailer you just signed, or the price change landing next month. Those go on top of the statistical baseline as explicit, separate adjustments, so you can tell afterwards which part of the miss was the model and which part was you. Promotions in particular need their true lift and their post-promotion dip, not just the spike — adjusting a forecast for promotions covers the mechanics.
5. Net it against what you already own
A forecast on its own is not an order. Subtract what is on hand, what is already on order, and what is committed to somebody else. What remains, dated against your lead time, is the buy.
6. Compare it to what happened
Then change something. A forecast that is never compared to actuals cannot improve, and the comparison is only worth anything if you keep the level and the period fixed while you do it.
How do you know whether a forecast is any good?
Start with the plainest measure there is: for each SKU, actual units divided by forecast units, expressed as a variance above or below. Above means you sold more than you planned; below means you sold less. Do it per SKU, per period, at the level you buy at.
Two things are worth separating once you have that:
- Error is how far off you were in either direction. It tells you how much buffer the SKU needs.
- Bias is whether you are consistently off in the same direction. A SKU that comes in above forecast eleven periods out of twelve does not have a variance problem, it has a broken assumption, and no amount of safety stock fixes an assumption.
Bias is the more actionable of the two and the more commonly ignored. Error gets averaged into a company-level figure that looks tolerable; bias hides inside that average and keeps costing money in the same direction every period.
Why this page does not quote an accuracy benchmark
You will find "good" forecast accuracy ranges quoted freely on the internet, usually without a source, and often traceable to another article quoting a third. We are not repeating them here. Accuracy depends so heavily on aggregation level, category and horizon that a single number is close to meaningless — the same catalogue can look excellent at total-company level and alarming at SKU-week level, with nothing having changed but the arithmetic.
The number worth tracking is your own, measured the same way every period, at the level you place orders at. Improving your own trend is a real goal. Hitting somebody else's unsourced benchmark is not.
What makes demand forecasting harder?
New products
No history means no statistical forecast. Use an analogous product you already sell, keep the first order deliberately small, and replace the assumption with real sales as fast as you can — forecasting demand for new products is the longer version.
Promotions and price changes
A promotion moves demand as well as creating it. Some of the spike is genuinely incremental and some of it is next month's sales pulled forward, and treating the whole spike as new demand is how brands end up long on inventory the month after a good promotion.
Multiple channels
DTC, Amazon, retail and wholesale have different demand shapes, different lead times, and different consequences for being wrong. Retail in particular is lumpy: an order for a launch is not a run rate, and treating it as one distorts everything after it.
Supplier constraints
The forecast says what you need. The supplier's minimum order quantity, case pack and container size say what you can actually buy, and the gap between those two is a decision about carrying cost rather than a forecasting problem. What MOQ means and how to decide on one covers that trade directly.
Signals the history cannot contain
Competitor moves, weather, an ingredient shortage, a platform changing its algorithm. This is the boundary between forecasting and demand sensing, which uses shorter-term signals to adjust a plan the statistical model has already produced.
Where Planster fits, and where it doesn't
Planster pulls sales and inventory from 150+ systems, tests several statistical models — Prophet, ETS, SARIMA and linear regression — against each SKU's own history, and keeps the one that fits, re-selecting it as the pattern changes. It detects seasonality at SKU level, lets you exclude stockout periods so they stop poisoning the fit, and builds one rolling forecast across DTC, Amazon, retail and wholesale. That forecast is then netted against what is on hand, on order and already committed, and turned into a reorder point, an order quantity and an order-by date per SKU. That is the engine, and everything else stands on it.
On top of it, the overnight run re-checks every SKU, order and channel and brings you a ranked list in the morning with the purchase orders already drafted. You open Planster, change what you want changed, and approve. Nothing reaches a supplier until you do. It is flat $1,000/month, and it is built for consumable CPG brands between $10M and $50M — food, beverage, supplements, beauty, household goods.
Where it does not help: if you have a handful of SKUs, one channel and one supplier, the arithmetic on this page is genuinely a spreadsheet job, and automating it will not be what changes your results. And no forecasting engine, ours included, produces a number for a product with no sales history — that first order is still a judgement call. See the forecasting engine for what the model selection looks like in the product, the agent for what arrives each morning, and pricing for the whole number on one page.