Services
Forecasting and demand modelling
Demand you can buy against. We build backtested time-series and demand models on your sales history, measure their error against a real baseline so you know how far to trust them, and wire them into the buying, replenishment or trading cadence your teams already run. Seasonality and promotional periods are modelled, not guessed, and every forecast carries a confidence range the buying team can read. A forecast only earns its keep when someone places an order on it, so the integration is the deliverable, not an afterthought.
The process we optimise
The process we optimise: forecasting demand accurately enough to buy and stock against it
Forecasting matters because someone has to commit to a buy, a stock level or a replenishment order before demand arrives, and the cost of getting that commitment wrong runs in both directions: capital tied up in stock that will not sell, and lost sales on the lines that do. So the process we optimise is not the forecast in the abstract, it is the buying and replenishment cadence the forecast has to feed. That means three disciplines most spreadsheet forecasts skip: measuring how wrong today's forecast actually is, so there is a baseline to beat; modelling the parts of demand that move it hardest, seasonality and promotional or markdown periods; and putting the forecast, with an honest read on its confidence, in front of the buying team at the moment they order. We build backtested demand models on your data and wire them into that cadence, then track forecast error over time so the number keeps earning the buyers' trust.
Before and after
What changes when the system is rebuilt
Before: a forecast no one has measured
- The forecast lives in a spreadsheet owned by one person, and stalls when they are away.
- No one has scored its error against actual sales, so the buying team cannot tell how far to trust it.
- Seasonality and promotional periods are handled by gut adjustment, and get worse the further ahead you look.
- The number sits beside the buying meeting rather than inside it, so orders are placed on instinct.
After: a measured forecast wired into buying
- A backtested demand model built on your data, with its error measured against your own history.
- Seasonality decomposed per line and market, and promotional or markdown periods modelled rather than guessed.
- The forecast reaches the buying and replenishment cadence with a confidence range the team can read at a glance.
- Forecast error is monitored after go-live, with a retraining plan, so accuracy is defended rather than assumed.
How we think about it
We start from the objective, then work outwards
The same discipline runs through every engagement: understand what the process is for, then design the system to serve it and measure against it.
Measure the baseline error before building anything
We begin by scoring the current forecast against what actually sold, because a forecast no one has measured is one no one can improve or fully trust. That baseline sets the bar the new model has to clear, and it is often the first time the buying team sees, in one number, how much the spreadsheet is costing them. For a private-equity-backed orthopaedics and medical-device supplier, an explicit baseline was what let the work be judged honestly, on a 31 per cent reduction in absolute forecast error with the horizon extended to 18 months.
Model what actually moves demand: seasonality and promotions
We decompose demand into the signals that drive it: the growth trend, the seasonality that is usually different for every line and market, and the promotional or markdown periods a naive model treats as noise. These are where spreadsheet forecasts fail hardest and where the money is. On general merchandise for a Big-4 UK supermarket, modelling how markdown depth and duration moved sales was what turned a reactive clearance process into a quantified £18m revenue-uplift opportunity.
Wire the forecast into the buying and replenishment cadence
A better number changes nothing if it lands after the order is placed or in a tool no one opens. We build the forecast into the cadence your teams already run, so it arrives in the buying, replenishment or trading meeting in a form they can act on. For a premium global footwear and lifestyle brand, machine-learning demand forecasts fed straight into a purchasing and reallocation plan across roughly 1.3 million items, recovering $3.8m of net revenue over six months by matching stock to where demand actually was.
Give buyers a confidence range, then monitor error over time
We hand the buying team not just a point forecast but an honest confidence range, so they can see how far to trust each number rather than treating them all alike. After go-live we monitor forecast error and agree a retraining plan, because a model that decays unnoticed quietly loses the buyers' trust. Packaging the logic into a lightweight, owned tool, as we did for the orthopaedics supplier, is what keeps the forecast in use once we have left.
An engagement, step by step
A representative forecasting engagement, drawn from how we structure this work:
- Weeks 1 to 3
Diagnose the buying cycle and measure the baseline
We map the buying and replenishment cadence the forecast has to feed, the sales history available, and how today's spreadsheet forecast is produced. We score its error against actual sales so there is a measured baseline, and agree the horizon and accuracy the buying decision actually needs.
- Weeks 3 to 5
Prototype and backtest the demand model
We build a first model that decomposes demand into trend, seasonality and promotional periods, and backtest it against your history so its accuracy is measured, not claimed. Backtesting early tells us whether the approach clears the baseline before we invest in productionising it.
- Weeks 5 to 8
Wire it into the ordering cadence
We harden the model and integrate it into the buying, replenishment or trading meeting, with a confidence range the team can read. The integration is the substance of the work: the forecast has to be present where the order is placed, not adjacent to it.
- Weeks 8 to 10
Monitor, retrain and hand over
We put forecast-error monitoring in place, agree a retraining plan, and hand over a documented, owned tool so the forecast is no longer one person's spreadsheet. We check that the buys it feeds are measurably better against the baseline we set at the start.
The engagement ends with a measured demand model wired into the buying cadence and monitored for error, not a notebook handed over. Buyers can see each forecast's confidence, retraining is planned, and the forecast lives inside the ordering cycle it was built to serve.
What you get
- Backtested demand models with error measured against a baseline
- Seasonality and promotional periods modelled, not adjusted by hand
- Integration into the buying and replenishment cadence you already run
- Forecast-error monitoring, a retraining plan and handover to your team
How it typically runs
Typical engagement: 6 to 10 weeks, 1 to 2 engineers, fixed-scope pilot.
Diagnose
Weeks 1 to 3
Prototype
Weeks 3 to 5
Build and integrate
Weeks 5 to 8
Productionise and hand over
Weeks 8 to 10
Indicative timeline. Every project is scoped individually: book a discovery call and we will provide a detailed proposal within 48 hours.
Where it fits
This practice is delivered through one of the two QuantSpark engines.
Proof