One of One Brands sells customizable, high-end apparel and accessories. Average order value is $250. Every order is different. Lead times vary depending on what was ordered. And their forecasting model was built for a world where historical sales predict future demand — which is exactly the assumption that breaks when every order is a custom configuration.
At $1.2M monthly revenue, they're not a small operation. But the operational complexity of custom products scales faster than revenue does. The forecasting problems that a standard apparel brand might avoid for years show up much earlier when the catalog is effectively infinite.
Why standard models fail for custom products
Most demand forecasting works by looking at what sold in a comparable past period and projecting it forward with adjustments for trend and seasonality. This works reasonably well when SKUs are stable and the demand pattern for each one is relatively consistent.
Custom products break both of those assumptions. The "SKU" for a custom item is effectively unique to each order. You can't look at historical sales of a specific configuration and extrapolate, because that exact configuration may never have been ordered before. And even if you define the SKU as the base product, the specific components ordered for each job vary enough that historical sales don't tell you what to have ready next month.
Stockouts happen on specific components even when overall demand looks healthy. Lead time variability compounds it, because a production run for a custom item can take weeks, so by the time you've identified what's running low, you're already behind.
Forecasting at the component level
The workaround that actually functions is to stop forecasting at the finished-product or SKU level and start forecasting at the component level.
For One of One Brands, this means looking at the underlying materials and options that appear across orders, not the specific finished products those orders produce. Some components appear in many configurations. Some materials are used across many SKUs. Those are the items that need safety stock, reorder points, and monitoring.
This requires mapping your product catalog to its components explicitly. In practice, most custom brands have more structured underlying components than they think. The infinite variety at the finished-product level is usually assembled from a finite set of materials, hardware, or base configurations.
Once you have that mapping, you're no longer trying to forecast which configurations will be ordered. You're forecasting demand for the underlying components, which have stable enough usage rates to be meaningful.
Using AOV trends as a demand proxy
One of the few consistent signals in a custom product business is average order value. When AOV trends up over several weeks, something is shifting — more orders, or higher-value configurations, or both. It doesn't tell you what components to stock more of, but it tells you whether overall demand is expanding or contracting. If AOV has been climbing for six weeks and production lead times are four weeks, that's context for a reorder decision even without a precise SKU-level forecast.
Treat it as a directional indicator for calibrating safety margins, not a number to put into a formula.
Building in larger safety margins
The model has to run with more buffer than the raw numbers suggest.
Standard safety stock formulas calculate the buffer you need to cover lead time variability given a normal demand distribution. For custom products, that calculation understates the risk on two dimensions. First, demand variability is higher than for standard products because customer preferences shift unpredictably. Second, production lead times are longer and more variable than for off-the-shelf inventory, meaning the cost of running out is higher because recovery takes longer.
The practical answer is to add a judgment-based margin on top of whatever the formula produces. For a brand where one production run takes four to six weeks, I'd rather carry an extra four weeks of component safety stock than be right on the math and wrong on the outcome.
With $1.2M monthly revenue and a 68% customer retention rate, One of One Brands has customers who come back. A stockout or extended delay on a custom order for a returning customer is a different kind of problem than losing a one-time sale. The safety margin calculation should account for what a missed order actually costs in customer lifetime value terms, not just the immediate lost revenue.
What this requires operationally
Component-level forecasting requires a bill of materials, or at minimum a structured mapping of configurations to components. If that doesn't exist, building it is the first step. Without it you're guessing.
It also requires accepting that the forecast will be directional, not precise. Custom product businesses are running with more uncertainty than businesses with stable SKU catalogs. The goal of forecasting in this context is to reduce the frequency of component stockouts and expensive emergency orders, not to achieve a specific fill rate against a number the model generated. That distinction matters for how you communicate accuracy to stakeholders who expect a standard forecasting model to perform like one.
