When the overstock report lands, the conversation is always about people. The buyers were too optimistic. The plan was too aggressive. Someone fell in love with a range. Next season we will be more disciplined. The report gets circulated, a promise gets made, and everyone moves on believing the problem was character.
Discipline is not the problem, and it is not the fix. Brands over-buy because the tools they plan with force a choice between two bad options: buy deep enough to avoid stockouts on the styles that work, or buy lean enough to avoid overstock on the styles that do not. With a single blended forecast, you cannot do both, so buyers hedge, and hedging at scale looks exactly like over-buying.
Once you see over-buying as a forecasting failure rather than a discipline failure, it stops being a lecture topic and becomes something you can measure and fix. You can point at the two mechanisms that produce it, quantify what each one costs, and change the inputs so the hedge is no longer rational. The buyers were never the problem. The forecast they were handed was.
A blended forecast hides the two mistakes that matter
One average number buries the size, the channel, and the style-level truth.
Most forecasts are built at a level of aggregation that is comfortable to reason about and useless to buy from. A category grows 6 percent, so the buy grows 6 percent. But you do not order a category. You order a specific style, in a specific size curve, for a specific channel, on a specific lead time, and none of those decisions can be made from a category-level number.
Underneath that tidy 6 percent, the medium sold out in week two and the extra-large never moved, the website ran hot while wholesale lagged, and two styles carried the whole category while four were dead on arrival. The blended number is right on average and wrong everywhere it matters, so the buyer pads the order to cover the parts they cannot see. The padding is not optimism. It is a rational response to being handed a number that hides the risk.
Take a worked example. A category is planned to grow 6 percent on a hundred thousand units, so the plan says buy a hundred and six thousand. Underneath, one hero style is pacing to double while three slow styles are pacing to half. A buyer who can only see the category number cannot fund the hero without over-funding the slow styles, because the blended plan raises everything by the same 6 percent. So they hedge: they buy the hero a little deeper to be safe and they do not cut the slow styles enough, because the plan gives them no license to. The hero still stocks out in week three, the slow styles still end at markdown, and the overstock report blames the buyer for both, when the real cause was a plan that could only move in one direction at once.
improvement in forecast accuracy (wMAPE) on hero SKUs when the model runs at SKU by size by channel instead of a blended category baseline.
Accuracy at the level you actually buy is what removes the need to hedge. When the forecast reads the size curve the merchandisers already see in real life, funds the hero and cuts the slow styles independently, and shows the channel split, the padding comes out of the order, because the order finally matches the demand instead of insuring against a blur.
Over-buying is the tax you pay for a slow signal
The plan is set once a season. Demand changes every day.
The second driver is cadence. A forecast that refreshes once a week, or once a season, forces every decision to carry a safety margin, because the next chance to correct is far away. The further apart your correction points, the more you buy to bridge the gap, because every order has to survive until the next time you can adjust it.
Shorten the interval and the math changes. A forecast that re-runs every hour against live sell-through, returns and channel mix means the next correction is never more than an hour away, so no single order has to hedge against three months of uncertainty. You buy closer to demand because you can adjust closer to demand. The safety margin shrinks not because the buyer got braver, but because the cost of being slightly wrong got smaller.
Return to the worked example and add cadence to it. If the buyer can only revisit that category at the seasonal plan, the hero has to be bought deep enough to last the whole season on one guess, and the slow styles have to be bought cautiously for the same reason, which is a hedge in both directions at once. If the buyer can revisit it hourly, the opening buy can be lean, and the hero gets its depth through reorders placed as the signal confirms it, inside the lead time. Same styles, same demand, far less capital committed to a guess. The slow signal was quietly financing the over-buy the whole time.
The overstock is concentrated, not spread
A handful of styles and sizes carry most of the excess. A blended plan cannot see them coming; a SKU-level one can.
Ten points of inventory is a lot of cash to leave in stock that will only ever clear at markdown. It is also the cash you need for the next buy, which is how over-buying becomes self-perpetuating: last season's overstock funds this season's caution, and the caution guarantees next season's stockouts, which produce next season's hedge.
Plan at the level you buy, correct at the speed demand moves
Then the discipline conversation takes care of itself.
Give buyers a forecast at SKU by size by channel that re-forecasts live, with a confidence range on every prediction, and the behaviour changes on its own. New styles inherit a size curve from their nearest historical cluster instead of a manual guess, so day-one launches ship with a real forecast. The buyer sees the range, not just the point, so they know exactly where the risk is and stop insuring against the rest. The hedge disappears because the reason for it disappeared.
The confidence range matters more than it sounds. A single point forecast forces the buyer to make one bet and pad it, because a point with no error bar is a number you either trust completely or hedge completely, and no one trusts a point completely. A forecast that says this style will sell between eight and twelve thousand units, with most of the probability near ten, tells the buyer exactly how much uncertainty there is and where. On the styles with a tight range they buy close to the forecast because the risk is small. On the styles with a wide range they open lean and let reorders fill in as the range narrows with real sell-through. The padding stops being a blanket applied to everything and becomes a precise response to where the actual uncertainty is, which is the difference between insuring the whole buy and insuring only the parts that need it.
In the worked example, that is the difference between committing a hundred and six thousand units to a plan that can only move one way, and committing a lean opening buy that funds the hero through reorders and never over-buys the slow styles at all. The overstock report the following season is shorter not because anyone was disciplined, but because the forecast finally let the buyer make the two decisions separately, at the speed the demand was actually moving.
Over-buying is not a character flaw. It is what rational people do when the forecast is too coarse and too slow to trust.
Fix the forecast and you do not need to lecture anyone about discipline. The buy comes down because the reason to hedge is gone, and the overstock report stops being a monthly search for someone to blame.