Solution 05 / Forecasting & Decision Models

Forecasts that change what someone actually does.

Demand and capacity forecasting, optimisation and scenario models, built around the decision they feed rather than the accuracy score they report.

Discuss Forecasting ↗

01The problem

Most forecasting effort is spent improving a number that nobody acts on differently. Accuracy is tracked while bias goes unmeasured, so the forecast is wrong in the same direction for months without anyone noticing. Worse, the model produces a single point estimate when the decision it feeds needs a distribution: how much buffer to hold, how much capacity to book, how much risk to carry. A mean is the wrong input for a decision whose costs are asymmetric.

Typical symptoms

Forecast accuracy reported monthly, with no record of any decision it changed
Bias never measured, so the same directional error repeats for months
A single point forecast used for decisions that need a range
Manual overrides applied routinely, with no check on whether they help or hurt
No naive baseline, so nobody can say whether the model beats last week repeated
Backtests run on random splits, quietly leaking future information into the past
Intermittent and slow-moving items forced through the same method as steady ones

Decisions we help you make

Which decision the forecast exists to serve, and at what horizon and granularity
Point estimate or full distribution, given how asymmetric the cost of being wrong is
Which series are forecastable statistically and which need human judgement
What method fits each demand pattern: smooth, erratic, intermittent, or lumpy
Where human override adds value and where it destroys it
How forecasts reconcile across a hierarchy of product, location, and period
What the decision rule is that converts a forecast into an action

02How we work on it

Methods we apply.

Decision-first framing: define the action and its cost asymmetry before choosing a loss function

Naive and seasonal-naive baselines, so every model must beat the trivial answer

Demand classification and forecastability scoring before method selection

Probabilistic forecasting: quantiles and prediction intervals, scored with pinball loss and interval coverage

Rolling-origin backtesting that respects time order and prevents leakage

Bias and Forecast Value Add analysis across each step of the process, including overrides

Hierarchical forecasting with reconciliation across product, location and time

Optimisation and scenario modelling where the decision is constrained, not just predicted

When being short costs more than being long, the decision is taken from a quantile, not the average. A model that only reports a mean cannot answer the question the business is actually asking.

03What moves

Metrics this work is measured on.

WAPE / MAPE against baselineForecast biasForecast Value AddPrediction interval coverageOverride rate and override valueThe downstream decision metric: service level, cost, or utilisation

04What we need to start

History by series and period, ideally 24 months or more
The decision this forecast feeds, and who makes it
Known drivers: promotions, pricing, events, capacity constraints
Current forecast and override history, if the two can be separated
What it costs to be wrong in each direction

05Engagement path

How this becomes an engagement.

Forecast Review

2 weeks

Benchmark current accuracy and bias against naive baselines, and establish whether the forecast is changing any decision.

Forecasting Build

6–14 weeks

Build the pipeline, method selection, backtesting harness and decision rule, then run it against held-out periods.

Monitoring Handover

Ongoing

Bias, drift and value-add monitoring your team owns, with retrain and escalation thresholds already tuned.

Want to see what this looks like against your own systems?

Most engagements start with a short, fixed-scope assessment, enough to quantify the opportunity before anyone commits to a build.

Start a conversation ↗