Three years of forecasting that leadership actually used

A US-market industrial manufacturer, roughly $70M in annual revenue, part of the Mueller group. DataWise built, delivered, and ran its sales forecasting for three years, inside a monthly cycle that fed directly into leadership planning. This page tells that story the way we graded the work itself: openly, against what actually happened.

What problem was the company trying to solve?#

Sales planning for the US market ran on experience and gut feel. Monthly and yearly targets, inventory commitments, and strategic planning had no quantitative forecast behind them. When customer behavior shifted, including during high-uncertainty periods, the shifts were detected late. Customer attrition was discovered only after the fact, when the orders had already stopped.

Nothing about this is unusual. It is how most mid-size manufacturers plan. The company’s leadership decided it should not stay that way.

What did the engagement actually look like?#

This was a three-year embedded data-science service, not a one-off model drop.

DataWise worked on a monthly delivery cycle: each month, updated forecasts and analyses were produced, graded against the previous months’ actuals, and presented into leadership planning meetings with the CEO and sales leadership. The forecast was not a report that landed in an inbox. It was a standing input to the meetings where targets, inventory commitments, and strategy were set.

That cadence is the real product. A forecast that leadership sees once is trivia. A forecast that leadership confronts every month, with its own track record attached, becomes part of how the company is run.

What did DataWise build?#

The core was multi-horizon sales forecasting, monthly and annual, built on an ensemble time-series stack that combined statistical models, gradient-boosted models, and neural models, tuned separately per horizon. A single model rarely wins at every distance; an ensemble tuned per horizon lets each approach carry the range it is best at.

Around that core, the engagement grew into a full analytical layer, all of it shipped and used:

• A yearly forecast performance ledger, updated monthly, tracking 3-month and 12-month forecast error against actuals. Predictions were graded, not just made.
• Customer-behavior analytics for periods of demand uncertainty: a live read on changing consumption patterns and buying habits while conditions were still moving.
• Churn analysis that matched detected attrition patterns against field intelligence from the sales organization, so leadership could explain a loss and act on it, not just count it.
• Customer clustering, to see the customer base in segments rather than as one undifferentiated total.
• Dedicated forecasting for mega-customers whose behavior was large enough to move the whole picture on its own.
• Stockpiling detection, separating real demand from customers buying ahead.
• An explainability layer, so forecasts came with reasons, not just numbers.
• Geographic sales analytics across the market.

How were the forecasts graded?#

This is the part we consider the centerpiece, and the part most forecasting vendors skip.

Every month, for three years, the forecasts were scored against what actually happened. The results lived in a yearly performance ledger with error tracked at the 3-month and 12-month horizons, and that ledger was on the table in the same leadership meetings the forecasts fed. When the model was right, everyone saw it. When it was wrong, everyone saw that too, and the next month’s cycle had to answer for it.

The page’s ledger visual shows the structure of that grading process. To protect client confidentiality, the figures shown are illustrative; the structure, the cadence, and the 3-month and 12-month error tracking are exactly what ran.

We hold ourselves to a simple standard: a forecast you never grade is an opinion.

What was the outcome?#

For three years, a company that had planned its US market on gut feel had a quantitative forecast, graded monthly against reality, as a standing input to its strategic planning. Shifts in customer behavior surfaced while they were happening. Attrition became something to explain and act on rather than discover.

We publish no invented performance numbers, here or anywhere on this site. What we state is factual: a three-year data-science engagement for a Mueller-group industrial manufacturer in the US market, with the CEO’s recommendation on file.

What would this look like for your company?#

You do not need a three-year commitment to find out. The starting point is a scoping conversation: where your planning decisions are made today, what data already exists around them, and whether a graded forecasting cycle would change what leadership sees each month.

If the answer is yes, we build it the same way we built this one: delivered monthly, graded openly, and answerable to your actuals.