Time Series Forecasting
Forecasting for the operational metrics that are not demand: capacity, traffic, energy, cash, headcount and volumes, backtested the way the future actually arrives and delivered with usable uncertainty.
Forecasting has a humbling literature. In repeated large-scale competitions, simple statistical methods have performed remarkably close to elaborate ones, and combinations of simple methods have often beaten both. Any engagement that starts by assuming a sophisticated model is required is starting from the wrong place.
Time series forecasting is the estimation of future values of a metric observed over time, using its own history along with seasonality, calendar effects and related drivers, evaluated by rolling-origin backtesting at the horizon that matters and delivered with prediction intervals rather than a single number.
The naive baseline is not a formality
The first thing we measure is what a naive forecast achieves: last period's value, the same period last year, or a simple moving average. This is not a courtesy. For many operational series it is genuinely hard to beat, and knowing the margin tells you whether the project is worth continuing.
We then measure your current forecast, whatever produces it today, because that is the real benchmark. Beating a naive method while losing to the planner's spreadsheet is not a result worth deploying.
Backtest by rolling origin, always
A single train and test split on time series tells you almost nothing, because it measures one arbitrary period. Rolling-origin evaluation refits and forecasts repeatedly across the history at your real horizon, which is both more honest and considerably less flattering. It is the first thing we build.
What a serious forecasting engagement covers
The horizon that matters
Accuracy at one step ahead is irrelevant if you make decisions six weeks out. We evaluate at the horizon your decision actually needs, and where several horizons matter, at each of them, because the best method frequently differs between them.
Calendar and known drivers
Holidays by region, working days, paydays, school terms, weather where it genuinely matters, and any planned events already in your calendar. Known future information is the cheapest accuracy available and is routinely left out.
Multiple methods, then combine
Classical statistical models, gradient boosting on lagged features and, where the data is plentiful, global models trained across many series. Combinations of several methods regularly outperform the best single one, which is one of the more reliable findings in the field.
Prediction intervals you can plan with
Buffers, capacity and risk limits need the spread, not the mean. Intervals are validated for coverage, because an interval that claims ninety per cent and achieves sixty is worse than no interval at all, and this is rarely checked.
Hierarchy and consistency
Where the series roll up, by site, region or business line, forecasts are reconciled so they sum correctly. Unreconciled numbers are how two teams end up planning against different totals derived from the same data.
Change points and regime shifts
Series break: a site opens, a policy changes, a pandemic happens. We detect structural breaks, handle the affected history explicitly rather than letting it pollute the model, and set an alert for when the pattern shifts again.
Where this applies
| Series | Decision it supports | Typical horizon |
|---|---|---|
| Contact and ticket volume | Staffing and shift planning | Daily to interval, 2 to 8 weeks |
| Energy consumption or generation | Procurement and balancing | Half-hourly to daily |
| Cash flow and receipts | Treasury and working capital | Weekly, 1 to 6 months |
| Infrastructure and capacity load | Scaling and provisioning | Hourly to weekly |
| Headcount and attrition | Recruitment planning | Monthly, 3 to 12 months |
| Claims or application volume | Operational capacity | Weekly to monthly |
Where the series is product demand, the specific structure of promotions, hierarchy and intermittency makes it a distinct engagement: demand and sales forecasting.
How the engagement runs
Naive and current baselines first, rolling-origin backtesting throughout.
Series audit and baselines
History assessed for gaps, breaks and outliers; naive and current-process baselines measured at your horizon.
Backtesting framework
Rolling-origin evaluation built at the horizons that matter, with the error measures agreed with the people who plan from it.
Model development
Statistical, machine learning and combined approaches, calendar and known drivers included, hierarchy reconciled.
Intervals and change points
Prediction intervals validated for coverage, structural break handling and shift alerting configured.
Delivery and handover
Automated forecasts into the planning system, monitoring, and the retraining runbook.
What you receive
Forecasts with honest error bars and a backtest anyone can reproduce.
Forecasting pipeline
Automated at your cadence and horizon, delivered into the consuming system as code.
Backtest results
Rolling-origin accuracy against naive and current baselines, at every horizon that matters.
Prediction intervals
Validated for coverage, usable for buffers, capacity and risk limits.
Reconciled hierarchy
Forecasts that sum consistently across sites, regions or business lines.
Change point handling
Structural break detection, treatment of affected history and shift alerting.
Monitoring and retraining
Accuracy tracking against forecast, with the trigger and procedure for refitting.
Is this the right engagement?
Worth being direct. Time Series Forecasting is the wrong spend in some situations, and those are listed rather than buried.
Good fit if
- An operational decision depends on a forward view of a metric over time.
- You have consistent history at the granularity you plan at.
- Uncertainty matters, because buffers or capacity are set from it.
- The current forecast is manual or has never been backtested.
- There is a system the forecast can be delivered into.
Choose something else if
- History is too short or too disrupted to contain a usable pattern.
- The series is driven entirely by decisions you make rather than by patterns.
- The metric is product demand, which is demand forecasting.
- Nobody will change a plan based on the forecast.
Frequently asked questions
Marked up with FAQPage schema so these answers can surface directly in search results and inside AI assistant responses.
Do we need machine learning for time series forecasting?
Often not. Classical statistical methods perform remarkably well on operational series, and combinations of simple methods frequently beat elaborate single models. We measure the naive baseline first and add complexity only where the backtest shows it earning its place.
How far ahead can you forecast?
As far as the series contains a stable pattern, which varies enormously. What matters is accuracy at the horizon your decision needs, so we evaluate there rather than reporting a flattering one-step-ahead figure.
Why do prediction intervals matter?
Because most operational decisions are about buffers: how much stock, how many staff, how much capacity. The mean tells you nothing about how much room to leave. We validate intervals for coverage, since an interval that claims ninety per cent and delivers sixty is worse than none.
What happens when something changes structurally?
We detect the break, handle the affected history explicitly rather than letting it distort the model, and alert when the pattern shifts again. Series that have been through a genuine regime change need that history treated deliberately, not averaged through.
How is this different from demand forecasting?
Demand forecasting is a specific application with product and location hierarchies, promotions, cannibalisation and intermittent demand. This covers the general technique for any operational metric, such as call volume, energy, cash or capacity.
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.
Demand and Sales Forecasting
SKU and location level forecasts with promotions, seasonality and hierarchy handled, measured where error costs money.
Read more →Custom ML Model Development
A bespoke model for a problem no off-the-shelf product fits, benchmarked against a simple baseline and deployed properly.
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.