How a reseller replaced gut feel with 86% forecast accuracy

Ever bought Nike or Adidas gear on Amazon? Chances are it came from one of the platform’s top 100 resellers for those brands.

Every reseller at that scale loses money in the same two places. Stockouts on the products that were selling, and write-offs on the products that were not. Both trace back to a single decision made monthly, by hand, on historical sales and category feel.

We rebuilt that decision around a forecast. Buyers now size purchase orders against a projected demand range with a stated confidence level, and the products with no sales history at all, where the guessing used to be worst, get forecast too.

Here is how the system works, and what forecasting demand across 12 million products taught us.

The challenge

Purchasing ran on experience, category feel, and Excel formulas. Someone with a strong read on the category reviewed inventory each month and decided what to buy and how much. That carried the business for years, and it broke down in four specific places.

No forward-looking signal
Reviews looked backwards at historical sales and nothing else. There was no systematic way to answer one question up front: is this specific product going to sell, and roughly how much, over the next few months. Without that answer every buy is a judgment call, and judgment calls do not scale.

Capital trapped in products that never sold
Money goes into stock that does not move, which is cash tied up in inventory plus storage plus whatever the team could have bought instead. Eventually it is a write-off.

Winners caught late, or missed entirely
The failure most teams underweight. Products that sold out early because nobody spotted them in time, or because nobody had the confidence to put real money behind them. Margin left on the table is harder to see than stock sitting in a warehouse.

New products worst of all
Two years of sales history at least gives you something to extrapolate from. A product with none is pure instinct, and those are the buys where the stakes run highest, because nobody has learned anything about them yet.

What we built

The system sits between the reseller’s sales history and its purchase orders. Two data sources feed an ETL pipeline on Google Cloud, which lands everything in one clean feature layer. Products are grouped by how they behave in the market, each group gets its own forecasting model, and a confidence layer turns the raw prediction into a range a buyer can act on. A forecast API serves it to the buying team, and a drift monitor retrains the models before accuracy quietly decays. The order itself stays with the buyer.

The sequence those layers run in is the whole trick.

Two sources feed it. The first is the client’s own sales, units, and pricing, pulled directly from Amazon’s Selling Partner API. The second is large scale external Amazon market data from Keepa, which is what lets the system say something useful about a product the client has never sold.

Both APIs throttle hard, with tight rate limits and quotas the pipeline is constantly bumping into at this scale. A naive ingestion job either falls over or silently drops data, and incomplete data quietly poisons every forecast downstream. So a real chunk of the work went into ingestion that respects those limits, backs off, retries, checkpoints, and resumes without double counting. That reliability layer is invisible when it works and catastrophic when it does not.

Conclusion

The largest recurring capital decision in the business now runs on a number instead of an instinct. Buying went from "I think this will sell" to "the model projects this range over the next three months, at this confidence, and here is the basis for it." It works on both ends: less money trapped in products that do not move, and more of the genuine winners caught early. Scope has expanded since launch, which is usually the clearest signal something is earning its place.

Products do not all sell the same way, and the whole system is built around that. Detect the pattern a product follows, route it to the behavioural segment and model that fit, and everything downstream works. Get the pattern detection wrong and no amount of model tuning saves you.

Want a forecasting system built like this one? Get in touch.