Forecasting · new products

Forecasting a new SKU with no sales history: borrow a launch curve, then let real sales take over

A new product has no history to forecast from, but you still have to place a purchase order before it sells a single unit. This post shows how to borrow the first weeks from SKUs you already sell, size the first order against a low, middle and high case, and move the forecast onto the new SKU’s own sales as the weeks come in. Every SKU and number in the worked example is invented so the arithmetic can be checked.

The standard forecasting advice starts with “look at your sales history.” For a SKU that launches next month there is none, which is why SKU-level forecasting gets one line on the subject: use a similar SKU or a manual estimate until the new one builds its own data. That is the right answer. This post is that line worked through, from the first purchase order to the week you stop needing the analogs.

It is a different job from forecasting a promotion, where the SKU has history and you are adjusting it for an event, and from the six-step demand forecast, which assumes the history is already there.

Step 1 — Choose the analog SKUs

An analog is a product you already sell whose launch looks like the one you are planning. The closer it is, the more of its curve you can borrow. Check four things, in this order:

1. Same channel. A product that launched on the same marketplace tells you more about a launch there than your best Shopify seller does, even if the product is closer.
2. Similar price. A $19 item and a $59 item in the same category do not sell at the same pace.
3. Same category and buyer. The new colour of an existing product is the easiest case there is. A new product type for a new buyer is the hardest.
4. A similar launch. Same time of year if you can, and a similar amount of advertising behind it. A launch that ran with heavy ad spend is not a fair analog for one that will not; ad spend moves the forecast on its own.

Pick two or three. One analog gives you a single guess with no sense of how wrong it might be. Three give you a range, and the range is the useful part.

Then pull each analog’s weekly units for its first twelve weeks after launch, not its recent sales. A product that sells 40 a week today may have sold 12 in its first week. If you forecast the new SKU from the mature rate, the first order will be sized for sales the launch has not built yet.

Worked example (invented SKUs and sales)

A brand is launching a new kitchen product on a marketplace. It has three earlier launches in the same category and price band. Their first weeks looked like this:

The “settled rate” is the average of weeks 9 to 12, after each launch levelled off. All three took about two months to get there. That ramp is the thing a mature-rate forecast misses.

These three become the low, middle and high cases for the new SKU. You are not averaging them into one number. You are keeping all three, because the next step needs them.

Step 2 — Size the first order against all three cases

The first purchase order has to last until the second one can arrive. That date comes from your lead time, not from your forecast.

In the example, the supplier’s lead time is ten weeks from purchase order to sellable stock (what lead time actually covers matters here, because a quoted production time is shorter than the real figure). The brand plans to place the second order at the end of week 4, after a month of real sales. Ten weeks later puts it on the shelf at the start of week 15, so the first order has to cover weeks 1 to 14.

Run each case to week 14, with weeks 13 and 14 at its settled rate:

Now test three order sizes against all three outcomes. For a slow launch the question is how long the leftover stock takes to clear. For a fast one it is how many units short you fall before the second order arrives.

The table works from twelve-week totals, so it shows how many units short each order falls, not the exact week the shelf empties. That depends on the week-by-week curve.

Leftover stock is measured in weeks to clear at the slower rate, not in dollars. That is the honest unit: the stock is not lost, it is tied up. What a week of being tied up costs you depends on your carrying cost and, if it is in a marketplace warehouse, on storage fees and how long you can wait before pulling it out becomes the cheaper option.

Running out costs something different: the units you could have sold between the empty shelf and the second order’s arrival. The right first order is the one whose bad outcome you can live with. If running out costs you far more per unit than carrying does, order toward the high case. If cash is tight, or the product could still flop, order toward the middle and accept the risk of running out.

The first order is not a forecast. It is a bet sized to the outcome you can afford to get wrong.

Two practical limits narrow the choice. The supplier’s minimum order quantity may already be above the low case, which settles that end for you. Cartons round every option up, so round to the case pack before you compare.

Step 3 — Blend in the real sales as they arrive

At the end of week 4 the brand has its first real data, and it has to place the second order now, because the lead time does not move. Four weeks is not much data, and the decision cannot wait for more.

The new SKU sold 8, 13, 17 and 20 units: 58 in four weeks. Compare that with each case’s first four weeks:

It is running between the middle and high cases, about 21% ahead of the middle. The tempting move is to scale the whole middle curve up by 21%. The careful move is to believe that 21% only partly, because four weeks is a short sample, and believe it more each week. This assumes the new SKU is following roughly the same-shaped ramp as the analog, just at a different level.

One simple heuristic is to give the real sales a weight that grows with the weeks you have. It is a rule of thumb, not a standard statistical method:

weight on real sales = weeks observed ÷ (weeks observed + k)

Here k is a dial you set, not a figure anyone can look up. It is the number of weeks of real data at which you trust the new SKU and the analog equally. A small k trusts the new SKU quickly. A large k waits. With four weeks observed:

The adjustment is the weight times the 21% gap: at k = 4, half of 21% is about 10%, and 30.25 a week times 1.10 is about 33.4. Pick k before launch and write it down, so the week-4 numbers do not decide it for you. If the analogs are very close to the new SKU, a larger k is reasonable. If they are loose, a smaller one is.

What the update tells you about the first order

Rerun the middle curve scaled up 10%. The 331 units ordered for weeks 1 to 14 now run out before the second order arrives at the start of week 15, about 35 units short, which is roughly a week of sales. The time to see that is week 4, while the second order can still be sized and booked to close the gap. In week 12 it is too late.

The second order is then sized like any other reorder, using the blended rate as the demand input: lead time plus the time until the next order, plus safety stock that reflects how unsure you still are. Because forecast uncertainty is still high for a SKU this young, you may choose a more conservative buffer than for an established SKU. The reorder point calculator covers the mechanics.

Two things that quietly break the update

Weeks when the SKU was out of stock

If the new SKU ran out for three days in week 3, that week’s units measure your stock, not your demand. Either leave the week out, or scale it to the days it was in stock (units ÷ days in stock × 7) and mark it as estimated. Use that only as an estimate: a short in-stock window can exaggerate demand just as easily as a stockout can suppress it. Counting it as-is makes a fast launch look slow, which leads to a smaller second order and another stockout. On a marketplace, judge whether a day was in stock from sellable units, not the total: reserved units are not on the shelf.

A launch push that will not repeat

Launch weeks can carry a discount, extra ad spend or a burst of attention from your own list. If the new SKU had that and the analogs did not, part of its lead over the middle case is the push, not the product. Treat it like the lift in a promotion forecast: estimate it separately and do not let it carry into the settled rate.

When to stop using the analogs

Keep re-blending at every review. As the weeks build up, the weight on real sales rises on its own: at k = 4 it is 50% at week 4, 67% at week 8 and 75% at week 12. At some point the analogs are adding noise rather than information, and the SKU moves onto the same forecast as the rest of the catalog.

A reasonable point to switch is when the SKU has about two months of in-stock weeks since its sales levelled off. That is a judgement, not a rule, and seasonal products need longer, because two good months in autumn say little about February. Write down the week you switched and why. When you plan the next launch, this SKU’s first twelve weeks become one of the analogs.

If the SKU settles near the low case and stays there, that is also an answer. It goes through the same review as any slow SKU, including whether it should be reordered at all.

What to do this week, if you have a launch coming

1. List two or three analog SKUs and write down why each one qualifies: channel, price, category, launch conditions.
2. Pull their first twelve weeks of sales, week by week, from launch.
3. Count back from your lead time to find the week the second order must be placed, and how many weeks the first order has to cover.
4. Build the three-by-three table: three order sizes against three outcomes. Pick the order whose bad outcome you can live with.
5. Choose k and write it down before the first sale.