AI Demand Forecasting & Production Planning

The problem: planning by gut feel, paid for in lost capacity

For a manufacturer, the demand forecast is the document that quietly decides everything else. It sets how much raw material to buy, which lines to run, how to sequence the work, and how many people to roster. When that forecast is little more than an educated guess in a spreadsheet, every downstream decision inherits the error — and the cost compounds.

That was the situation a leading footwear manufacturer with €235M in turnover brought to us. Day-to-day production planning spanned a wide and seasonal SKU range, and it was largely manual. Planners leaned on experience, last year's numbers and a feel for the market. It worked, until it didn't: the business was over-producing some lines while running short on others, tying up cash in stock that moved slowly and missing sales on the styles customers actually wanted.

On top of the demand uncertainty sat a hard physical constraint. Every mould change on the line cost capacity — setup time, warm-up, scrap and lost throughput. Without a reliable demand signal to plan around, mould changes were reactive, triggered by surprises rather than scheduled into a sensible sequence. Each unnecessary or badly-timed change was capacity quietly leaking out of the factory. Meanwhile, aggressive competitors were happy to take any market share that slipped.

The brief was practical, not academic: give planners a demand signal they can trust, and turn that signal into a production plan that respects how the plant actually runs.

How the system works

We did not try to replace the planners. We built a forecasting and planning layer that does the heavy statistical lifting, then hands a clear, explainable recommendation to the people who know the floor. The pipeline runs in three stages.

1. Forecast — Prophet for seasonality and trend

The first job is to model the shape of demand. Footwear is deeply seasonal, with overlapping patterns: yearly seasons, holiday peaks, promotional spikes and slow long-term trends per style family. We used Prophet, an additive time-series model that decomposes a series into trend, seasonality and holiday effects. Prophet is a strong fit here because it handles multiple seasonalities cleanly, copes with missing data and irregular history, and produces a forecast a human can actually read — you can see the trend line and the seasonal curve, not just a number. That transparency matters when a planner has to defend a production decision.

2. Refine — gradient boosting (XGBoost) for the drivers

Seasonality and trend explain a lot, but not everything. Real demand also responds to price changes, promotions, weather, channel shifts and the long tail of SKU-specific quirks. To capture those, we layered a gradient-boosting model (XGBoost) on top. Where Prophet models the smooth, structural part of the series, XGBoost learns the residual patterns from a richer set of features and corrects the forecast as fresh data arrives. The combination is deliberate: Prophet keeps the forecast stable and interpretable, while gradient boosting sharpens accuracy on the messy, driver-heavy reality. Together they are more accurate than either model alone, and they degrade gracefully — if a SKU has thin history, the structural model still gives a sensible baseline.

3. Plan — turning the forecast into a schedule

A forecast that sits in a dashboard changes nothing. The third stage turns the demand signal into production scheduling. Forecasts drive the build plan so that runs are sized to real expected demand, and — critically — so that mould changes line up with genuine need instead of reacting to surprises. By sequencing production around a trustworthy forecast, the plant cuts the number of disruptive changeovers and the capacity each one wastes. The result is a plan that is both demand-aware and physically realistic for the line.

Data and approach

Good forecasting is mostly good data engineering. Before any model can earn its keep, the history has to be clean, aligned and trustworthy — which is why we treat the data foundation as the first deliverable, not an afterthought. For this engagement we consolidated historical sales and order data across the SKU range, reconciled calendars and promotional events, and built a repeatable pipeline so the models always train on consistent, current data.

• Backtesting before go-live. We validated the models against held-out historical periods, comparing forecast accuracy to the manufacturer's existing planning baseline so the uplift was measured, not assumed.
• Per-SKU and aggregate views. Forecasts are produced at the level planning actually needs — granular enough to schedule a line, aggregated enough to plan capacity and procurement.
• Continuous correction. The gradient-boosting layer re-corrects on fresh data, so the forecast tracks reality instead of drifting between manual updates.
• Human in the loop. Planners review and can override the recommendation; the system supports their judgement rather than overruling it.

This is the same disciplined machine-learning approach we apply across our work: pick the simplest models that fit the problem, validate them honestly on real data, and wrap them in tooling people will actually use.

Results — real, delivered outcomes

These are real results from a Crux Digits engagement; the client name is withheld pending permission to publish it. Within a year of putting the forecasting and planning system to work, the manufacturer saw:

• +8% revenue growth within a year, as better demand alignment meant the right styles were available when customers wanted them.
• +7% production-capacity improvement, recovered largely by smoothing mould-change losses and running the lines against a realistic plan.
• Market position retained at #4 against heavy, aggressive competition — defending share that had been slipping.

The pattern is the one that makes forecasting worth doing: it pays back on both sides of the ledger at once. Revenue rises because you stop missing demand, and cost falls because you stop wasting capacity. The forecast becomes the single source of truth that aligns sales, procurement and the factory floor.

Who this is for, and the ROI case

This approach fits any business where demand is uneven and supply is constrained — most obviously across manufacturing, but equally in retail and wholesale distribution, food and consumer goods, and anywhere planners juggle a large, seasonal product range. If you recognise any of the following, there is almost certainly value to capture:

• Planning runs on spreadsheets and experience, and accuracy swings with whoever is doing it.
• You routinely carry too much slow stock while stocking out of the fast movers.
• Changeovers, setups or batch switches eat a meaningful share of capacity.
• Forecasts are updated occasionally by hand, so they are stale by the time decisions are made.

The ROI maths is usually straightforward. A few points of forecast accuracy translate directly into less working capital tied up in inventory, fewer lost sales, and recovered production hours — and those gains recur every planning cycle. Because the system builds on tools and data you already have, the cost to stand it up is modest against the return. We keep pricing transparent and project-based; you can see how we structure engagements on our pricing page.

How we would run a pilot for you

We do not ask you to take forecasting accuracy on faith. We prove it on your own data first. A typical pilot follows a short, low-risk path:

• Scope and data check. We agree the SKUs and horizon that matter most, then assess the history you have and what it can support.
• Baseline and backtest. We measure your current forecasting accuracy, then build and backtest the Prophet-plus-gradient-boosting models against it — so the improvement is a number, not a promise.
• Forecast and plan. We deliver demand forecasts feeding into a production plan, with the human-in-the-loop review your planners need.
• Decide with evidence. You see the measured uplift on your own operation before committing to a full rollout.

That is exactly how we delivered the result above, and it is how we de-risk every engagement. We are an Utrecht-based AI consultancy and software studio working across the Netherlands, the Benelux and Europe — and turning models into production systems is the part we are built for. If forecasting and planning is a live problem for you, this kind of work sits squarely within our AI implementation practice. For a related example of cutting waste on the factory floor, see our predictive maintenance case study, or for the demand-side equivalent in consumer goods, our churn prediction case study. When you are ready to scope a pilot on your own numbers, get in touch.