AI Stock Reordering When You Barely Have a Sales History

The message from the supplier says the fabric you use is going up eight percent next month and asks whether you want to order now. You have sold, in the last year, somewhere between eleven and nineteen of the thing you make from it, depending on whether you count the market stall. You have no idea whether to buy for six months or eighteen.

This is the reorder decision as it actually presents itself to a small seller, and it is not the problem the forecasting industry describes. That industry assumes thousands of SKUs and years of daily data. You have forty lines, two years of patchy history and a supplier whose lead time is whatever it turns out to be.

What follows is the version of the problem that fits, and an honest account of where a model helps.

Why does ordinary forecasting fail on a small catalogue?

Because most of your lines sell nothing most weeks, and that breaks the assumptions underneath almost every forecasting method.

The technical name is intermittent demand. A study of slow moving item modelling defines it simply as demand per unit of time being zero in some periods, and notes that traditional approaches in this space lack the properties of a statistical distribution, which is why they perform poorly when plugged into inventory models. Their own pharmacy example ran at 30 percent of periods with zero demand. A small maker's catalogue is often far above that.

The classic answer, Croston's method, works by splitting the problem in two: how big the demand is when it happens, and how long the gaps between demands are. That split is genuinely clever and it is the right mental model even if you never implement it. Your slow line does not sell 0.3 units a week. It sells three units, roughly every ten weeks, and those are different facts with different consequences for how much you hold.

The practical implication is one most small sellers get wrong. An average that includes the zero weeks will tell you to hold almost nothing, and then you will be out of stock every time the lumpy demand arrives. We covered how to choose a forecasting method and evaluate it honestly for anyone who wants the deeper version; this piece stays with the decision rather than the model.

What decides how much cover to hold?

Not your nerves, though that is what usually decides it. There is a formula and it is one line, and it is worth the five minutes because it converts a feeling into a number you can defend.

The newsvendor model asks how much to order when demand is uncertain and you get one shot. Stitch Fix's explanation states the optimal condition as F(Q*) equal to the underage cost divided by the sum of underage and overage cost, where the underage cost is the cost of being one unit short, typically your margin, and the overage cost is the cost of an unsold unit, typically original cost less salvage value or the cost of carrying it another period.

Work an example with your own numbers and the answer often surprises. Say you make a candle for four pounds and sell it for sixteen. Being one short costs you twelve pounds of margin. An unsold one, if you can eventually sell it at a market for eight, costs you four in tied up value at worst. The critical ratio is twelve divided by sixteen, or 0.75, meaning you should hold enough to cover demand three quarters of the time. That is a lot more cover than most makers carry, and the reason is that your margin is high and your salvage value is decent.

Now run the same calculation for something perishable or seasonal, where the unsold unit is worth nothing after December. Suddenly the overage cost is the full production cost, the ratio drops, and the correct answer is to hold less and accept some stockouts. Same business, opposite instruction, and the only thing that changed is what an unsold unit is worth.

How much safety stock, with two years of thin data?

Here the answer is refreshingly blunt: use the crude method. The average to maximum method, stated as maximum daily demand times maximum lead time minus average daily demand times average lead time, is the one recommended for companies with minimal historical information, on the grounds that it gets you a working number fast.

It is crude because it keys off your worst observed case rather than off a distribution, and with few observations your worst case is unstable. It is also honest about that. The statistical alternatives need a standard deviation you cannot estimate reliably from nineteen sales.

When you do have enough history, the service level approach becomes available, and it is worth understanding what it costs. The same source gives the mapping: 90 percent service level corresponds to a Z value of 1.28, 95 percent to 1.65, 97.5 percent to 1.96 and 99 percent to 2.33. A worked example elsewhere shows the arithmetic in practice, with a 141.4 standard deviation in demand, a 1.15 month lead time and Z of 1.28 producing 194 units of safety stock.

Read those Z values as a price list. Moving from 90 to 99 percent service does not cost ten percent more stock, it costs roughly 82 percent more safety stock, because the multiplier goes from 1.28 to 2.33. For a small business with cash tied up in inventory, choosing 99 percent across the board because it sounds responsible is one of the more expensive defaults available.

The variable nobody measures

Lead time. Ask a small seller how long their main supplier takes and you get a number from memory that is usually the best case they once experienced.

This matters more than the forecast because lead time sits inside every formula above, often twice: once as the period you need to cover and once as a source of variance. The combined formula quoted by the same source, Z times the square root of average lead time times demand variance plus average demand squared times lead time variance, makes the point visible. Lead time variability is multiplied by average demand squared. A supplier who is sometimes three weeks and sometimes nine weeks will dominate your buffer requirement no matter how stable your sales are.

The fix is a spreadsheet with four columns: order date, promised date, arrival date, quantity received against quantity ordered. Six months of that is worth more than any forecasting tool, and it also gives you the evidence for a conversation with the supplier, which is frequently the cheapest way to shrink the buffer.

So where does AI actually come in?

Three places, none of them the one in the advertising.

The first is data cleaning, which is the unglamorous reason most small forecasting attempts fail. Your sales history is split across a till, a website and a market stall spreadsheet, with product names that do not match. Getting those into one series with consistent identifiers is a text matching job, it is tedious, and it is exactly what a model is good at. Nothing downstream works until this is done, and most people give up here.

The second is the calculation itself, used as a calculator rather than an oracle. Give it your numbers and ask it to compute the critical ratio, the reorder point and the buffer under each method in the table above, then show its working. You are using it to avoid arithmetic mistakes, which is legitimate, and you can check the result because you supplied the inputs.

The third is the review. Once a quarter, ask it to list every line where cover exceeds six months of sales, every line with no sale in ninety days, and every line where you have ordered more than twice in the last quarter, which usually means the reorder point is set too low. That is pattern spotting across a table, and it surfaces the decisions worth making.

What it should not do is produce a demand forecast from nineteen data points and present it as a number. It will do this, confidently, and the output will look like the outputs it has seen. On a thin series that confidence is not knowledge, and acting on it with cash is how a small business ends up with a stockroom full of something that sold well once.

The ABC question, done in twenty minutes

One classification is worth doing before any of this, because it stops you applying the same rule to a line that earns half your income and a line that earns nothing.

Sort your products by annual revenue contribution. The top group, typically a handful of lines producing most of the money, deserves a high service level, a measured lead time and a real buffer. The middle group gets a simple rule and a quarterly glance. The bottom group, which in a small catalogue is often half the lines producing a tenth of the revenue, should mostly be made to order, held at one unit, or discontinued.

The guidance in the safety stock literature points the same way: set service targets from a classification rather than from an arbitrary percentage applied everywhere. For a small seller the practical version is that three different rules is plenty, and that applying one rule to everything is what causes both stockouts on the important lines and dead capital on the unimportant ones simultaneously.

Retiring a line is the decision small sellers avoid hardest and it is frequently the most profitable one available. Capital sitting in a slow line is capital not sitting in the thing that sells, and the carrying cost is real even when it is invisible.

Back to the eight percent price rise

The question that opened this piece now has a method rather than a feeling behind it. A supplier price increase is a one shot decision under uncertainty, which is the newsvendor problem in its original form, and the tempting answer, buy a lot because it is about to cost more, is right only under conditions you can check.

The saving is eight percent of the material cost, which on a four pound item is thirty two pence a unit. Against that sits the carrying cost of holding the extra stock for however long it takes to use, plus the risk that the design changes, plus the cash that is no longer available for anything else. Buying eighteen months of cover to save thirty two pence a unit is a poor trade for almost any small business, because the binding constraint is cash rather than cost of goods.

The version that usually works is to buy to your normal reorder quantity plus one cycle, which captures most of the saving on the stock you were certainly going to need anyway and commits nothing to demand you are guessing at. If the increase were thirty percent rather than eight, the arithmetic changes and a deeper buy becomes defensible. The point is that the threshold is calculable, and the supplier's deadline is not a reason to skip calculating it.

Running this where the stock actually lives

All of the above assumes one number for what you hold, and a surprising share of small businesses do not have one. Stock is in the shop, in a spare room, allocated to an unfulfilled order and listed on two marketplaces that each think they own it.

The reorder point is meaningless against an unreliable count, which is why the first step in the diagram above is a physical count rather than a query. If you are setting up or rebuilding the online side, having one stock record that the website, the till and the marketplace listings all read from removes most of the reconciliation work permanently, and it is much easier to do at the start than to retrofit. Our AI store builder treats inventory as a single record for that reason rather than as a sync problem to solve later.

The related habit is to reconcile deliveries as they arrive rather than at year end, which is the same discipline that keeps apparent stock loss from being blamed on theft when it was a counting error.

The forecast is the part everyone wants to improve. The lead time is the part that decides the answer, and almost nobody writes it down.

A starting sequence for this week

Count your top ten lines physically and compare against the system. Note the variance, because that number tells you how much to trust everything that follows. Open your emails and reconstruct the last six deliveries: what you ordered, when you ordered, when it arrived. That is your lead time baseline.

Then for each of those ten lines, write down the margin and what an unsold unit is worth to you in six months. Compute the critical ratio. For the lines where it comes out high, you are probably underordering. For the lines where it comes out low, you are probably holding too much.

Then set a reorder point per line as demand over the lead time plus the buffer from the average to maximum method, and put it in whatever system you already use. That is the whole job for a small catalogue, it takes an afternoon, and it will outperform any tool applied to an unreliable count.