What Data Do You Need to Predict Customer Reorders?

The short answer

You already have the data. Predicting when a recurring customer will reorder needs nothing more than the order history a distributor records as a matter of course. The useful fields are the account, the product, the quantity, and the date of each order.

This surprises people who assume prediction takes some exotic feed. It does not. The whole signal sits in the transactions a distributor has been logging for years, which is why a distributor can start from records they already own.

What each field contributes

Each piece of the record plays a part in finding the reorder pattern.

• Account: ties the orders together so each customer's rhythm is read on its own.
• Product: a customer reorders different items on different cycles, so the pattern is per product.
• Quantity: order size tracked against the account's own average, so a shrinking order is visible.
• Date: the spacing between order dates is the raw signal for cadence and for spotting a slip.

What you do not need

You do not need anything from inside the customer's building, you do not need them to tell you their schedule, and you do not need a new system of record. That is worth stating plainly, because it also marks the limit: this reads ordering behavior toward you, and nothing else. The order history in an Epicor P21 or Eclipse system, or even a clean export, holds enough to read the pattern.

Turning the data into a prediction

The method is arithmetic, not magic. Average the gaps between an account's past orders, add that average to the last order date, and you get an expected next order date with a window of slack around it. Orders placed within a few days of each other count as one, and it takes about four spaced orders before the average means anything. Do that for every account and product, count days against each window, and the accounts furthest past their own gap rise to the top of a ranked list.