The situation
The business runs a physical store and an online shop, with around forty SKUs that do the bulk of the volume. On paper that sounds manageable. In practice, the buyer was spending most of her week trying to second-guess what would sell next month, and getting it wrong often enough that it was starting to hurt.
The visible problem was a stockroom that had quietly turned into a graveyard. Slow movers had piled up over several seasons, taking up space, tying up cash, and slowly losing value as newer ranges came in. Some lines were being marked down twice just to clear them, and a few were eventually written off entirely. The ops lead had a rough sense of how much was sitting there, but nobody wanted to add it up properly because the number was uncomfortable.
The less visible problem, and arguably the more painful one, was what happened on the bestsellers. The lines that customers actually came in for kept running out. Repeat buyers would turn up, find an empty shelf or a sold-out page, and either wait, walk, or buy from someone else. The team could feel they were losing sales but had no clean way to measure how many.
The buyer was doing all of this by hand. She had a spreadsheet, a gut feel built up over years, and the supplier lead times memorised. What she did not have was time. Every reorder decision was a small judgement call made under pressure, and the pressure was constant. Cash flow was getting squeezed at the wrong end: too much money in stock that would not move, not enough product on the shelves that would.
What we did
We started by pulling two years of sell-through data out of the till system and the website, and lining it up properly. Not glamorous work, but it had never been done in one place before. Once we could see how each SKU had actually behaved week by week, the patterns the buyer half-knew became visible and measurable.
From there we built a forecasting model on top of that history. Nothing exotic. It looked at recent sell-through, longer-term trend, and seasonality, and produced a weekly demand estimate for each line along with a recommended reorder quantity that respected supplier lead times and minimum order sizes. We layered in adjustments for the obvious seasonal swings the business already knew about, and a few less obvious ones the data flagged that the team had not consciously noticed.
The important design choice was leaving the buyer in the loop. The model produced recommendations, not orders. Every Monday she sat down with the week's suggested purchase list, looked at anything unusual, and either approved it, edited it, or overrode it entirely. If a supplier had told her something the model could not know, she changed the number. If a promotion was coming, she factored it in. The point was never to take her judgement out of the process. The point was to stop her judgement being burned on routine reorders so she could spend it where it mattered.
We also gave her a simple view of what was sitting in the stockroom and how long it had been there, so the slow movers had nowhere to hide.
The result
After the first full quarter, waste on slow-moving stock was down by eighteen percent against the same period the year before. Availability on the top selling lines improved by thirty-one percent, measured as the share of weeks each bestseller was in stock and ready to sell.
The number that mattered most to the owner was not on the report. It was the cash that came unstuck. Money that had been sitting in dead inventory started flowing back through the business as those lines were cleared and not replaced. New ranges could be funded out of that recovered cash rather than out of the overdraft. The buyer got a chunk of her week back, and now spends it talking to suppliers and looking at new ranges instead of firefighting reorders.
It is not a finished piece of work. The model gets retrained as new data comes in, and the buyer keeps finding edge cases worth feeding back. But the business is now organised around what customers are actually buying, rather than what someone hoped they would buy three months ago.