Demand Forecasting & Planning

Most demand-forecasting systems start at too high a level — by category, by region, by week — because that is where the data is cleanest. The decisions that benefit from forecasting, however, happen at a much finer grain: which SKU to ship to which store on which day, how much labour to schedule for which shift, how much capacity to reserve in which warehouse next month. We build forecasts at the operational level of the decision, not the level of the data, and we use hierarchical reconciliation to ensure that the lower-level forecasts roll up coherently to the higher levels the business reports on.

Working at SKU-store-day, SKU-channel-week or equivalent fine-grained levels brings its own challenges: long-tailed demand, intermittent series, and series with very short history. We use specialised methods (Croston-style intermittent forecasters, hierarchical pooling, embedding-based new-product methods) to keep accuracy acceptable on the long tail rather than abandoning it.

Hierarchical reconciliation — using methods such as MinT and trace-minimisation — lets us produce forecasts that are coherent across the hierarchy and almost always more accurate than independently-produced forecasts at any single level. The discipline is straightforward and the gain is consistent.

• SKU-store-day, SKU-channel-week and similar grains
• Intermittent and slow-moving series
• New-product and short-history forecasting
• Hierarchical reconciliation (MinT, top-down, bottom-up)
• Calendar, promo, weather, macro signals
• Probabilistic forecasts for safety-stock policy

Demand is shaped by far more than its own history. Promotions, price changes, weather, holidays, school calendars, sporting events, macroeconomic indicators and competitor actions all leave fingerprints on demand patterns. We incorporate driver data into the forecasting model where it has been shown — by careful out-of-sample testing — to improve accuracy more than it adds operational complexity. We are equally happy to leave drivers out where they do not pay back; the value is in the discipline of measurement, not in the inclusion itself.

Promotional effects receive particular attention. Promotion data is notoriously messy, with overlapping mechanics, varying execution quality and inconsistent post-event measurement. We work with marketing and trading teams to clean and standardise promotional history before feeding it to the model, and we explicitly model uplifts, halo effects on related products, cannibalisation of substitutes, and post-promotion dips. The accuracy gains during promotional periods are usually substantial.

Weather and calendar effects are added with similar care, with appropriate weighting for the categories and locations where they matter. A frozen-food retailer in a northern climate has very different weather sensitivity from a beverage business in a warm one; the model captures both.

“A forecast that ignores the promotional calendar is a forecast that mostly fits the past.”

A demand forecast that lives in a separate dashboard rarely changes a planning decision. We integrate forecasts directly into the planning systems where decisions are made — Anaplan, Kinaxis, o9, SAP IBP, Blue Yonder, JDA, bespoke planning tools — so that the forecast becomes the default starting point that planners refine, not a parallel artefact they consult and ignore.

Override workflows are built in deliberately. Planners hold context the model does not, and a system that cannot accept their input will quickly be sidelined. We design overrides to capture the reason alongside the value, so that the override pattern itself becomes a learning signal: which planners override the model in which categories, with what justification, and with what subsequent accuracy.

We also pay close attention to the consensus process. The S&OP cycle reconciles demand, supply and finance views into a single agreed plan; a model that throws fresh numbers into that process without respecting its rhythm will be more disruptive than helpful. We design the cadence of the model output, the level at which it enters the consensus discussion, and the comparison to the previous cycle, so that the model strengthens the process rather than competing with it.

Forecast accuracy is a means; service level and working capital are the ends. Our deployments are explicit about how an improvement in accuracy will be converted into operational benefit: tighter safety-stock policies derived from probabilistic forecasts, earlier reorder points, smarter allocation across stores, smoother production schedules. Without that translation, accuracy improvements stay academic.

Probabilistic forecasting is central. By producing the full predictive distribution rather than a single point, we let inventory policy be set against a target service level explicitly: a 95% in-stock target requires more buffer than a 90% target, and the probabilistic forecast tells you exactly how much. The same principle drives capacity reservation, supplier commitments and labour scheduling.

We pair the forecast with a clearly-defined measurement programme. Pre-deployment baselines on service, stockout rate, days of cover and write-off are established explicitly; post-deployment performance is tracked against the same metrics, on a comparable basis, by quarter. The intent is that the value of the system can be defended in a finance review, not just demonstrated in a model dashboard.

Engagements normally run in three phases. A six-to-eight-week diagnostic and prototype establishes the baseline, characterises the data, builds a candidate model, and quantifies the expected operational benefit. A three-to-six-month build phase productionises the model, integrates it into the planning environment, and runs in parallel with the existing process. A subsequent operate phase covers monitoring, retraining, ongoing improvement and the gradual extension to additional categories or geographies.

We are equally happy to deliver this with a fully embedded team or to provide a small specialist group that supports your existing planning and IT functions. The shape of the engagement is dictated by the way you want to operate the system afterwards.