Demand Forecasting for a New Waterproof Bag Launch: Plan the Range, Not the Number

A new product forecast is a structured guess, and should be written as one

The instinct on a launch is to produce a number: four thousand units, because the spreadsheet says so. That number then becomes the order, the cash commitment and the plan, and when reality differs the difference is treated as a failure of forecasting rather than as the expected behaviour of an estimate. The more useful habit is to treat the new product demand forecast as a distribution with a low case, a base case and a high case, and to make the purchasing decision from the shape of that distribution rather than from its centre. Where forecast accuracy measurement earns its keep is not in grading the guess afterwards, but in calibrating how wide the band should have been in the first place.

Practically, this means three changes to how the number is written. Every launch forecast states a range with the reasoning behind each boundary. Every launch order is split into a committed quantity and a reserved quantity, where the reserve is an intention to reorder rather than an order. And every launch has a written decision rule — at week six, sell-through above X means reorder, below Y means stop — agreed before the goods arrive so that the decision is not made under pressure.

The reason this matters more in this category than in many is the replenishment structure. With sampling taking 6–10 working days, bulk production 35–50 days and transit beyond that, a reorder placed after launch lands somewhere between twelve and sixteen weeks later. If the season is twenty weeks long, the reorder may arrive after two thirds of the demand has already happened. That is not a reason to over-order; it is a reason to plan the reorder decision early and to accept that the first buy is the season for practical purposes.

It is also a reason to separate two questions that get conflated. How much will this product sell in its life, and how much must I buy before I know anything? The first is genuinely unknowable at launch and will be answered by the market. The second is a planning problem with a calculable answer, and it is the one worth spending the effort on.

Why new product forecasts are structurally wrong rather than merely imprecise

Forecasts for established products are imperfect in a manageable way: the error is mostly noise around a known level, and it shrinks as more history accumulates. New product forecasts fail differently. The error is structural, it is biased in a predictable direction, and it does not shrink with effort because the underlying data does not exist.

Two of those six deserve emphasis because they account for most of the systematic optimism. Analogue mismatch: people pick the analogue they want to be like, which is usually the best-performing comparable rather than the most similar one. And launch support variance: a forecast built assuming a front-page placement, an email to the full list and a trade-show reveal will not survive any one of those not happening.

The consequence is that a new product forecast should be deflated deliberately before it is used. A common working practice is to take the honest base case and treat it as the optimistic case, then plan the first buy at a figure below it. That is not pessimism; it is a correction for a known bias, and it is cheaper to be pleasantly surprised into a reorder than to be left holding the season.

Method one: the analogue product, and how to pick a genuinely comparable one

The analogue method is the default for a reason: it is fast, it is free, and it is better than nothing. Its entire quality depends on how the analogue is chosen, and the default choice — the product most like this one that sold well — is the one most likely to mislead. The discipline is to score candidates on several dimensions and to prefer similarity over performance.

The last row is the one most often skipped and it carries the largest correction. A new product has no reviews, no search history, no repeat buyers and no word of mouth; a comparable product measured at maturity has all of them. Applying the mature rate of sale to a new launch without a discount is one of the most reliable ways to over-order. A working correction is to take the analogue’s own launch-period rate rather than its steady-state rate, and then to apply a further discount for genuine novelty.

Where several analogues exist, use the spread rather than the average. If three comparable launches did eight hundred, one thousand four hundred and two thousand two hundred units in their first season, the useful output is not one thousand four hundred and sixty-seven — it is a range of eight hundred to two thousand two hundred, which tells you what the error looks like and therefore what the plan should be.

Method two: channel and customer research that actually predicts

Research can sharpen a forecast, but most of what is called research does not predict behaviour, it predicts opinions. The distinction is well established: stated purchase intent overstates actual purchase by a large multiple, and the gap widens when the question is hypothetical, the price is not stated, and the respondent has no commitment. The useful forms of research all involve either money or a real commitment.

• Pre-orders or a waitlist at the intended price, with a stated delivery date — the closest thing to real demand you can get before production.
• Retailer or distributor indications, collected as quantities per account and marked as non-binding, then discounted heavily.
• A landing page with real traffic and a real price, measuring conversion rate rather than clicks or sign-ups.
• A paid reservation or a small deposit, which converts enthusiasm into a number you can plan against.
• Comparison against a previous launch where you ran the same research, so you can calibrate the conversion from stated intent to actual sales.
• Avoid: surveys asking whether someone would buy, in the abstract, at an unstated price, from a friendly audience.

Calibration is the part that makes research useful rather than decorative. If you ran a waitlist before a previous launch, you know what proportion of waitlist sign-ups converted to sales; that ratio is the single most valuable forecasting asset a small brand can own, because it turns an opinion into a quantity. Record it, and apply it to the next launch.

Retailer indications deserve a specific warning. An account saying "we would take five hundred" is not an order, and the gap between indication and purchase order is routinely wide — buyers are optimistic by disposition, and their own sell-through may not support the number when the time comes. Collect the indications, discount them, and weight the ones from accounts that have a reorder history with you far more heavily than the ones from new accounts.

Method three: the test sell, and what a valid test requires

A test sell is the only method that produces genuinely new information, because it exchanges a small amount of inventory for real data. It is also the easiest method to run badly, and a badly run test produces confident nonsense. Three conditions make a test valid: it runs at the intended price, in the intended season, with the intended channel and marketing support.

Price is the most frequently violated condition. A test run at an introductory discount measures demand at the discounted price, which tells you nothing about demand at full price and typically overstates it substantially. Season is the second: a waterproof bag tested in the wrong season measures the wrong thing entirely, because the category is strongly seasonal. And channel: a test sold through one route does not transfer cleanly to another, because the customers and the conversion mechanics differ.

• Quantity: enough to sell out or clearly not sell out across representative doors — often one to three hundred units. Below that, a single large order or a single quiet week dominates the result.
• Duration: four to eight weeks in season. Less than that measures launch curiosity rather than a sustained rate of sale.
• Measurement: units sold per door per week, and weeks of cover remaining. Totals hide velocity, and velocity is what you are ordering against.
• Conditions: full price, real marketing, normal placement. Any deviation changes the demand being measured.
• Comparison: a control product selling alongside it, which separates the new line from the effect of the season or the promotion.

The measurement unit matters as much as the design. Record the rate of sale per week per selling location, not the total, because the total cannot be extrapolated and the rate can. If a test sells sixty units in three weeks across four doors, the useful number is five units per door per week, which can be multiplied by the number of doors you intend to open at launch. A total of sixty tells you very little.

And decide in advance what the test decides. A test that ends without a decision rule has produced expense rather than information. Write the rule first: above a stated rate, launch at the high case; between two rates, launch at the base case; below a floor, do not launch or launch at the minimum only for a further read. That is the whole value of the exercise.

Blending the three methods into a range

The three methods produce three different kinds of output — an adjusted rate from an analogue, a calibrated quantity from research, and a measured velocity from a test — and the temptation is to average them. Do not. Average them and you get a spurious central number with an invisible error. Combine them as bounds and as weights instead.

• Treat the analogue spread as the outer bounds of plausible demand, because it is the widest and least precise input.
• Treat calibrated research as a check on the base case, and give it weight only where you have a historical conversion ratio to calibrate against.
• Treat a valid test as the strongest single input, and let it move the base case rather than merely inform it.
• Where the methods disagree widely, widen the range rather than resolving the disagreement — disagreement is information about uncertainty.
• State the assumptions explicitly next to the number: price, channel count, marketing support, season length.
• Re-issue the range whenever an assumption changes, because a forecast conditional on support that did not happen is simply wrong.

A worked shape helps. Analogue launches suggest eight hundred to two thousand units in a first season. A waitlist calibrated at a known conversion ratio suggests around one thousand two hundred. A test at five units per door per week across a planned forty doors for a sixteen-week season suggests about three thousand two hundred, and that last figure is the one to distrust until the door count is real. The honest output is not one number but a stated range with the reasoning attached, and a first buy set at the conservative end of it.

The reason to keep the reasoning attached is practical: three months later, when you are deciding whether to reorder, what you need is not the number but which assumption failed. If demand was low because the door count was lower than planned, the product may be fine and the channel plan wrong. If the doors opened and the rate was weak, the product or the price is the issue. Only a documented forecast lets you tell those apart.

Measuring forecast error: MAPE, bias and the direction that matters

Once the product has sold for a season, the forecast can be graded — and grading is only useful if it separates two different things: how far off you were, and in which direction you were consistently off. Those are different problems with different fixes, and the standard metrics exist precisely to separate them.

Bias is the more actionable of the two, because a consistent bias can be corrected while random error cannot. If your last four launches over-forecast by an average of thirty per cent, then every future launch forecast should be deflated by roughly that amount before it becomes an order. That single correction is worth more than any improvement in method, and it costs nothing once you have recorded the data.

Direction also matters asymmetrically for inventory. Over-forecasting produces dead stock that is slow and visible; under-forecasting produces stockouts that are fast and largely invisible, because the lost sale never appears anywhere. Businesses systematically over-forecast — surveys of launch performance consistently find optimism bias — and the practical response is to deflate the first buy and reserve the ability to reorder rather than to try to guess better.

Keep the record in a single file: product, launch date, forecast range, actual sales, error, bias, and one line on why. Ten launches of that data turns forecasting from an argument into a calibrated process, and it is the only forecasting asset that compounds. The formal definitions are standard; the Institute for Operations Research and the Management Sciences publishes reference material on forecasting methods and error measures if you want the textbook versions.

Forecast error multiplied by lead time determines the buffer

This is the single most useful relationship in launch planning, and it is arithmetically simple. The inventory you need as protection is a function of how wrong the forecast might be and how long it takes to correct the error. Get either one wrong and the buffer is wrong; get both roughly right and the buffer is close enough to be useful.

Take a concrete case. Planned demand is two hundred units a month. Total replenishment lead time — sampling, bulk production at 35–50 days, transit and receiving — is around twelve weeks, so lead-time demand is roughly five hundred and sixty units. If the forecast error measured on comparable launches is plus or minus thirty-five per cent, the plausible swing over a lead time is about two hundred units either way. That two hundred is the scale of the buffer worth holding, and it is worth comparing directly against the holding cost of those units.

Notice the square-root relationship in the fourth row. Buffering against variability grows with the square root of lead time, not with lead time itself, which is why shortening lead time is powerful: halving a twelve-week lead time to six weeks reduces the variability component by about thirty per cent, not by fifty. It is still the single most effective structural improvement available, and it is why committing to a reorder window early in the season matters more than refining the forecast.

The comparison at the end of the table is the decision. Two hundred units held at a carrying cost of roughly twenty per cent of a twelve dollar landed cost is about four hundred and eighty dollars a year, against a stockout of a hundred units at eighteen dollars of margin, which is one thousand eight hundred dollars. The buffer is cheap and the stockout is expensive, and that relationship is stable across most of this category. Our safety stock and buffer strategy guide sets out the fuller calculation, and the inventory turnover guide puts the carrying-cost side of it in context.

Writing the forecast as a range with a commit quantity and a reserve

The output of all this should be one page with three numbers rather than one. A commit quantity, which is what you order now. A reserve, which is the additional quantity you intend to order if demand justifies it, with the capacity and timing pre-arranged rather than hoped for. And a stop condition, which is the sell-through rate below which the line is not reordered.

Commit-and-reserve planning is standard practice rather than a novelty, and the Association for Supply Chain Management publishes reference material on the demand-planning process if you want the formal framework behind it. The commit quantity should sit at or below the conservative end of the range, for a reason that is arithmetically straightforward. The cost of the low outcome is a stockout at the top of the range, which is expensive per event but bounded and recoverable through a reorder if the timing allows. The cost of the high outcome is inventory that does not sell in the season and carries into the next one, where it occupies capital and space and is eventually written down. These are not symmetric, and the asymmetry favours buying the low end.

The reserve only works if it is real. That means the production slot, the material position and the reorder decision date are agreed before the first order ships, because a reorder discovered at week eight and placed at week ten arrives after the season. Practically, ask for the reorder window in writing when you place the first order: what quantity, at what lead time, if ordered by what date.

The stop condition is the part people omit, and its absence is why weak lines get a second buy. Write it before launch: if sell-through at the read point is below a stated percentage of the low case, the line is not reordered, regardless of how good it looks in isolation. Pre-committing to that rule removes the sunk-cost argument at exactly the moment it is most persuasive.

Launch quantity against minimum order quantity: the constraint that shapes the plan

Minimum order quantity is the constraint that most often overrides a good forecast, and it should be designed around rather than absorbed. With a minimum of 500 pieces per style, the first buy cannot be smaller than that, which means a launch whose low case is three hundred units carries a structural risk of two hundred units of dead stock before anything is known.

• Launch one style rather than five: concentrate the first buy in a single style and colourway. It costs a narrower offer at launch, and the rest of the range follows in season two.
• Use one colourway for the first buy: variants can be added on the reorder once the base colour has proven the rate of sale.
• Treat the minimum as the test: accept 500 as the test quantity and plan the read point from it, accepting higher initial exposure than a smaller test would need.
• Share the buy with a channel partner: commit volumes against an account indication to share the risk, at the cost of margin and dependency.
• Accept a higher unit cost on a smaller run where that option exists: it is frequently worth paying against dead-stock risk.

The first row is the most reliably correct. Variant proliferation is the single most common reason a launch over-orders: five colourways at five hundred pieces each is two thousand five hundred units against demand that may be eight hundred, and the split between the variants is itself unknown. Launching one well-chosen colourway converts an unknowable five-way split into a single measurable rate.

There is also a sequencing argument. The most efficient use of a first season is to learn the rate of sale for the product, not to maximise the first season’s revenue. A launch that sells out at the minimum and generates a confident reorder is a better outcome than one that sells seventy per cent of a large buy and produces an ambiguous signal, because the second one leaves you unsure whether the product or the quantity was wrong.

For smaller programmes, the trade-offs are laid out in our small-batch guide for startups, and the sampling stage that precedes any of this is described in the sampling guide. Both sit upstream of the quantity decision and both change what the minimum actually means.

The first season: what to read, and when

Once the product is selling, the forecast stops mattering and the rate of sale takes over. Two metrics carry almost all of the information: units sold per door per week, and weeks of cover remaining at the current rate. Everything else — revenue, conversion rate, page views — is downstream of these or too noisy to plan on.

• Set the read point before launch: commonly weeks four to six of the season, early enough that a reorder can still land inside the season.
• Compare the actual rate against the rate implied by the low, base and high cases, not against a single number.
• Compute weeks of cover remaining, and compare it against total replenishment lead time plus buffer weeks.
• If cover is less than lead time plus buffer, reorder immediately — the decision does not improve with waiting.
• Separate the door count question from the rate question: low total sales with a healthy per-door rate is a distribution problem, not a demand problem.
• Track the returns rate from the first month; a launch returns rate materially above your norm changes both the forecast and the reorder.

The reorder trigger deserves to be mechanical, because the human version of it is unreliable. Define it as: weeks of cover remaining is less than lead time plus buffer weeks. If lead time is twelve weeks and buffer is five, then at seventeen weeks of cover you reorder, and not before. Running that rule automatically removes the two failure modes — reordering late because the stock looked adequate, and reordering early because sales felt strong.

Watch the returns signal carefully on a new line, because it is the fastest available indicator of a product problem rather than a demand problem. A returns rate well above the norm in the first weeks usually indicates a specification, sizing or expectation issue, and reordering into that is how a manageable problem becomes a large one. The analysis of warranty and return rates covers how to read that signal against your own baseline.

When to reorder, and when to stop

Two decisions end the season: reorder or not, and continue or discontinue. Both should be rule-based rather than editorial, because the emotional pull at that moment is towards the product you have invested in rather than towards the evidence in front of you.

Reorder when the rate supports it and the timing works. Concretely: the per-door rate is at or above the base case, weeks of cover are at or below the reorder trigger, and the reorder will land with enough season left to sell through. If the reorder would arrive with less than a full selling window remaining, the correct answer is usually to wait for the next season rather than to buy into the tail.

Stop when the rate does not support it, when the returns signal is poor, or when the line is cannibalising a stronger existing product. Cannibalisation is the subtlest of these because total revenue can look healthy while incremental profit is negative: if the new line is largely drawing sales from an existing product with better margin, the launch has cost money rather than made it. Measure the existing line’s rate before and after launch to see it.

And record the outcome against the forecast, in the same file, with one line on the cause. That is the only part of this process that makes the next launch better, and it is the part most often skipped because by then everybody has moved on.

A one-page forecast template you can reuse

Everything above fits on one page, and keeping it to one page is what makes it get used. The fields below are the minimum that supports a real decision; anything beyond them tends to be elaboration that delays the order rather than improving it.

• Product and launch date: style, colourway, variant count and the intended first selling week, which anchors everything else to a calendar.
• Assumed price and channel count: launch price, number of doors or channels, and expected support. The forecast is conditional on these, so state them where they can be checked.
• Analogue range: comparable launches with their first-season quantities and a similarity score, providing the outer bounds.
• Research estimate: a calibrated figure, with the conversion ratio used, providing a check on the base case.
• Test result: units per door per week, or "not run" — the strongest single input where it exists.
• Low, base and high: three numbers, each with one line of reasoning. This is the forecast itself.
• Commit quantity: what is ordered now, and why that figure rather than the base case.
• Reserve and reorder window: additional quantity, latest order date and expected arrival, which makes the reserve real rather than aspirational.
• Buffer and trigger: the weeks of cover at which you reorder, making the decision mechanical.
• Stop condition: the sell-through rate below which the line is not reordered, which removes the sunk-cost argument in advance.
• Outcome, filled in after the season: actual sales, error, bias and one line on cause — the only asset that compounds.

Fill it in for every launch, including small ones, and review three or four of them together once a year. The pattern across launches is far more informative than any single one: it tells you your bias, your realistic error band, and which of the three methods has actually worked for your business.

Programmes quote FOB Xiamen with a minimum of 500 pieces per style, sampling in 6–10 working days and bulk production in 35–50 days, which means the lead-time input in that template is known in advance rather than guessed — and the reorder window can therefore be agreed at the time of the first order. If you want that schedule documented for your own launch plan, review our process from first enquiry through sampling into bulk production and send us your specification, target volume and intended launch date.

Frequently Asked Questions

Q1. How accurate can a new product forecast realistically be?

Not very. Errors of 30–50% are normal for a genuinely new product even with a good process, and 40–60% is common. Plan the range and the reorder window rather than trying to sharpen the central number.

Q2. What is MAPE and what is a good score?

Mean absolute percentage error: the average size of the miss, ignoring direction. Below 20% is unusual for a launch; 30–50% is a normal working range.

Q3. Why does bias matter more than MAPE?

Because a consistent bias can be corrected. If your last four launches over-forecast by about thirty per cent, deflate every future launch by that amount — a correction that costs nothing.

Q4. What is the best method for forecasting a new bag?

A test sell, if it can be run at full price in the right season with real marketing support. Where a test is not possible, a scored analogue with a deliberate discount for novelty.

Q5. How do I pick a comparable product?

Score candidates on channel, price band, use case, season, launch support and variant count — and prefer similarity over performance. Using the best-selling comparable rather than the most similar one is a common error.

Q6. Do customer surveys help forecast demand?

Only in forms that involve commitment: pre-orders, deposits, or a landing page with a real price measuring conversion. Abstract purchase-intent questions overstate demand substantially.

Q7. How big should a test sell be?

Enough to sell out or clearly not sell out across representative doors, often one to three hundred units, run for four to eight weeks in season and measured as units per door per week.

Q8. How much buffer stock do I need for a launch?

Roughly the forecast error applied to lead-time demand. At two hundred units a month, twelve weeks of lead time and 35% error, that is about two hundred units — cheap compared with the margin lost to one stockout.

Q9. Does shortening lead time reduce the buffer a lot?

Meaningfully. The variability component grows with the square root of lead time, so halving a twelve-week lead time cuts that component by roughly thirty per cent.

Q10. Should I order the base case for the first buy?

Usually the conservative end of the range instead. The downside of under-buying is a bounded stockout that may be recoverable; the downside of over-buying is stock that carries into next season.

Q11. How does a 500-piece minimum change the plan?

It sets a floor on the first buy. Launch one style in one colourway rather than five variants, so the minimum tests one measurable rate instead of an unknowable split.

Q12. When should I decide whether to reorder?

At a read point set before launch, commonly weeks four to six — early enough that a reorder can still land inside the season. Decide on weeks of cover against lead time, not on sentiment.

Q13. What is a reorder trigger I can apply mechanically?

Reorder when weeks of cover remaining falls below total lead time plus buffer weeks. With twelve weeks of lead time and five weeks of buffer, that is at seventeen weeks of cover.

Q14. How do I know if a new line is cannibalising an existing one?

Measure the existing product’s rate of sale before and after launch. Total revenue can look healthy while incremental profit is negative if the new line draws from a higher-margin product.

Q15. What should I do if returns are unusually high after launch?

Treat it as a product signal rather than a demand signal, and do not reorder into it. A launch returns rate well above your baseline usually indicates a specification or expectation problem.

Q16. Is selling out at the minimum a failure?

No. A launch that sells out at the minimum and generates a confident reorder is better than one that sells seventy per cent of a large buy, because the second leaves you unsure whether the product or the quantity was wrong.

Q17. What should I record after each launch?

Forecast range, actual sales, error, bias and one line on the cause. Ten launches of that data turns forecasting from an argument into a calibrated process.

People Also Ask

How do you forecast demand for a new product?

Use the spread of comparable launches for the bounds, calibrated pre-orders or waitlists as a check, and a full-price in-season test sell as the strongest input — then plan a range, not a number.

What is a good MAPE for new product forecasting?

Thirty to fifty per cent is a normal working range for a genuinely new product, and 40–60% is common. Below twenty per cent is unusual.

How much safety stock do I need for a launch?

Approximately the forecast error applied to lead-time demand: at 200 units a month, twelve weeks of lead time and 35% error, around 200 units or five weeks of cover.

Should I order more or less than the forecast?

Usually less. The cost of under-buying is a bounded stockout, while over-buying leaves stock that carries into next season and is eventually written down.

When should I reorder a new product?

At a pre-set read point, triggered when weeks of cover fall below total lead time plus buffer weeks — early enough that the reorder lands inside the season.

What is a test sell?

A small launch quantity sold at full price in the right season with normal marketing, measured as units per door per week, to produce a rate you can extrapolate.