Demand and Sales Forecasting
Forecasts at the level you actually order and staff at, with promotions, seasonality, holidays and price changes handled explicitly, and accuracy measured where the error costs money rather than on a comfortable aggregate.
Every business already has a demand forecast. It is usually a spreadsheet, last year plus a percentage, adjusted by whoever knows the account. That forecast is the benchmark, and beating it is less automatic than vendors suggest, particularly on the slow-moving items where most of your inventory cost is sitting.
Demand and sales forecasting is the production of quantity forecasts at the level decisions are made, typically product and location by week or day, accounting for seasonality, promotions, price, holidays and lifecycle, reconciled across the product and location hierarchy, and evaluated on the horizon and granularity that drives ordering and staffing.
Why aggregate accuracy flatters everyone
A total forecast for the business next month can be very accurate while every individual decision it informs is wrong. Errors at the item level cancel out in aggregate, and the aggregate is what usually gets reported.
We measure where the money is: at the product and location level, at the horizon that matches your ordering lead time, weighted by value. That number is always worse than the aggregate one, and it is the only one that predicts whether stock will be in the right place.
Intermittent demand needs a different method
Slow-moving items with many zero periods break conventional forecasting, and averaging produces a fractional forecast that is never right. These need methods built for intermittency and, crucially, a different accuracy measure. In most catalogues the long tail is where the excess stock and the write-offs actually accumulate.
What a serious forecasting build includes
The calendar, properly
Public holidays by country and region, school terms, paydays, moving festivals, sporting fixtures, seasonal transitions and your own trading calendar. Getting the calendar right typically improves accuracy more than the choice of model, and it is routinely under-invested in.
Promotions and price
Uplift, cannibalisation of related products, pull-forward of future demand and the post-promotion dip, modelled explicitly rather than absorbed into a residual. Without this, a promoted week teaches the model the wrong lesson about baseline demand for months afterwards.
Hierarchical reconciliation
Forecasts produced at several levels and reconciled so that item forecasts sum to the category and the region totals the business plans with. Unreconciled forecasts are why planning and finance argue about which number is real.
New products and end of life
Items without history forecast from analogues and attributes, and discontinued items handled deliberately rather than forecast forward indefinitely. Every catalogue has both, and both are usually handled by someone overriding a spreadsheet.
Uncertainty, not just a number
Prediction intervals so safety stock and staffing buffers can be set from the actual spread rather than from a service-level rule of thumb. A point forecast alone tells a planner nothing about how much room to leave.
Backtesting the way the future arrives
Rolling-origin evaluation over multiple periods, at your real lead time, with no information from the future available to the model. This is where most vendor accuracy claims quietly fall apart, and it is the first thing we build.
What it is worth
| Improvement | Operational effect | Where it shows up |
|---|---|---|
| Fewer stockouts on fast movers | Sales recovered that were previously lost | Revenue and availability measures |
| Less excess on slow movers | Working capital released, fewer markdowns | Inventory value and write-off |
| Better staffing forecasts | Less overtime and fewer idle hours | Labour cost per unit |
| Fewer manual overrides | Planner time recovered | Planning cycle time |
| Honest uncertainty | Safety stock set to actual variability | Service level at lower stock |
We baseline each of these before deployment so the improvement is measured rather than claimed, and we report accuracy against your existing forecast rather than against a naive method that nobody uses.
How the engagement runs
Your current forecast is the benchmark, backtested the way the future actually arrives.
Data and baseline
History assembled with promotions, price and calendar, and your current forecast measured at item and location level as the benchmark.
Backtesting framework
Rolling-origin evaluation at your lead time and granularity, with value-weighted error measures agreed with planning.
Model development
Candidate models including intermittent-demand methods, promotions and price effects, hierarchical reconciliation.
New products and uncertainty
Analogue-based forecasting for items without history, lifecycle handling, prediction intervals for buffer setting.
Integration and handover
Forecasts delivered into the planning system with override tracking, plus monitoring and retraining guidance.
What you receive
Forecasts your planners will use, with the accuracy measured where it matters.
Forecasting pipeline
Automated at your cadence and granularity, delivered into the planning system as code.
Accuracy benchmark
Your current forecast against the new one, at item and location level, value weighted, over multiple backtest periods.
Promotion and price effects
Uplift, cannibalisation and post-promotion dip modelled and quantified.
Hierarchical reconciliation
Forecasts that sum consistently across product and location levels.
Prediction intervals
Uncertainty per item and horizon, usable for safety stock and staffing buffers.
Override tracking
Where planners adjust the forecast and whether those adjustments improve it, reported monthly.
Is this the right engagement?
Worth being direct. Demand and Sales Forecasting is the wrong spend in some situations, and those are listed rather than buried.
Good fit if
- You hold stock or staff to a forecast and errors cost real money.
- You have at least a couple of years of history at the level you plan.
- Promotions, price changes or strong seasonality affect demand.
- Planners currently override a spreadsheet forecast extensively.
- There is a planning system the forecast can be delivered into.
Choose something else if
- History is too short or the product mix changes completely every season.
- Demand is driven almost entirely by a handful of contracts rather than by patterns.
- Nobody will change ordering or staffing based on the output.
- The metric is not demand but another time series. See time series forecasting.
Frequently asked questions
Marked up with FAQPage schema so these answers can surface directly in search results and inside AI assistant responses.
How much history do we need for demand forecasting?
Two full seasonal cycles is the comfortable answer, and useful work is possible with less where the seasonality is weak or where similar products provide analogues. What matters more than length is whether promotions, price and calendar events are recorded, because without them the model learns the wrong pattern from the same history.
How much better will the forecast be than ours?
We will not quote a percentage before backtesting against your own forecast, because the honest answer depends on how good yours already is. Businesses with a heavily overridden spreadsheet usually see a large improvement; those with a mature statistical process see a smaller one concentrated in the long tail.
Can it forecast new products with no history?
Yes, from analogues and product attributes rather than from the item's own history, with wider uncertainty stated explicitly. This is one of the places where planners spend the most manual effort today, so it is usually worth including in scope.
Should we still let planners override the forecast?
Yes, and we track those overrides and report whether they improved accuracy. In most organisations some planners consistently add value with overrides and others consistently do not, and having that measured changes the conversation constructively.
How is this different from time series forecasting?
Demand forecasting is a specific application with its own structure: product and location hierarchies, promotions, intermittency and lifecycle. Time series forecasting covers the general technique for any metric over time, such as capacity, traffic or cash.
Often paired with this
Most clients combine two or three engagements from the Machine Learning & Predictive Analytics pillar. These are the ones that most often run immediately before or after.
Time Series Forecasting
Forecasts for any metric over time, backtested properly and delivered with usable uncertainty.
Read more →Predictive Analytics
Predictions embedded in the decisions they are meant to improve, measured on the business outcome rather than on AUC.
Read more →Optimisation and Operations Research
Scheduling, routing and allocation solved against your real constraints, with solutions planners will actually run.
Read more →Is this the right engagement?
Tell us what you are trying to build. If a different service fits better, or if you do not need us at all, we will say so.