FORECASTING & DEMAND PLANNING
Planning on a spreadsheet that has not been right since 2019
Capacity planning, raw material ordering and delivery commitments usually rest on a mixture of last year’s numbers and the planner’s judgement. That works until the mix shifts. The data that would do better — actual cycle times by product and machine, real changeover durations, genuine yield rather than standard yield — generally exists in the shop-floor systems but never reaches the planning spreadsheet in a usable form.
What the system does
Forecasting what the plant can do
Capacity forecasting using actual observed cycle times, changeovers and availability by machine and product — not routing standards that were set when the equipment was new.
Demand and consumption
Order and consumption history modelled with its seasonality and trend, so raw material and consumable ordering reflects the pattern rather than the last three months.
Ranges, not single numbers
Every forecast carries a prediction interval. A planner who knows a figure is “180 to 240, most likely 205” makes better commitments than one handed a confident 212 with no indication of how wrong it might be.
Scenario comparison
What happens to delivery dates if that machine goes down for three days, or if the mix shifts toward the slow product? Scenarios run against the same model so the comparison is consistent.
Scored against what happened
Forecast accuracy is tracked and published. A forecasting system whose error is never measured is a confidence machine, not a planning tool, and we would rather show you the error than hide it.
Grounded in the shop floor
Because the MES already holds genuine production history, forecasts rest on what the plant actually did rather than on what the ERP assumed it would do. This is the difference that makes the numbers usable.
How a forecasting project runs
Define the decision
Forecasting is only worth doing if a decision changes as a result. We start from the decision — how much to order, what to commit to, when to schedule maintenance — and work backwards to what needs predicting and how accurately.
History assessment
How much usable history exists, at what granularity, and with what gaps or regime changes. We will tell you if it is too thin to support the forecast you want, and what a simpler intermediate step would be.
Model and backtest
Models are tested against held-out history, so you can see how they would have performed over periods you remember. Backtesting against a year your planners lived through is the most convincing evidence there is.
Deliver into the planning process
Forecasts appear where planning actually happens, with intervals and scenario tools, and accuracy is reported every period. The model is retrained as the mix and the equipment change.
What this is built on
Production history from the MES database joined with order and consumption data from the ERP. Time-series modelling in Python — seasonal decomposition, gradient-boosted regression on engineered calendar and mix features, and classical statistical methods where the series is short and a simple model genuinely outperforms a complex one. Outputs delivered through the MES interface and as scheduled exports in whatever format your planning process already consumes.
What decision are you making with incomplete numbers?
Tell us the planning decision that is hardest to get right and what history you hold. We will tell you honestly whether a forecast would move the needle or whether better instrumentation comes first.