Forecasting uses patterns through time to estimate what may happen next. Trend, seasonality, promotions, shocks and volatility can each affect demand, while an annual average hides the variation managers must prepare for. Marketways separates these influences, tests forecasts on periods the model has not seen and reports a plausible range around the estimate. Clients can plan capacity, inventory, cash and staffing at the level where the fluctuations occur.
The decision this method supports
We use Forecasting & Time-Series Analysis to help clients answer: What range of future outcomes is supported by the available time-series evidence?
How the method works
Time-series analysis studies observations recorded in sequence. Forecasting uses recurring patterns, trend, seasonality and relevant external drivers to estimate future values. A sound forecast is tested against data the model did not use and compared with simple benchmarks.
A business example
A distributor may forecast monthly demand for each product. The analysis must distinguish genuine seasonality from promotions and stockouts. Forecast ranges can then show where inventory decisions remain exposed to uncertainty.
How the client uses the result
Forecasting helps a business prepare capacity, inventory, cash and staffing before demand is known. A range of plausible outcomes allows leaders to plan a central case while retaining options for stronger or weaker conditions.
What we deliver
We produce forecast ranges, scenarios, risk distributions, prioritised alternatives or an optimised plan. A manager should understand which assumptions drive the result and what conditions would trigger a different action.
Limits and complementary methods
A forecast helps predict what may happen, but it does not necessarily explain why. Structural change can make a previously accurate pattern unreliable.
Selected methods and techniques
We select from these established methods according to the decision, evidence and operating conditions.
- ARIMA / SARIMA: Apply the autoregressive integrated moving average (ARIMA) or seasonal ARIMA (SARIMA) model family to a defined time series. The analyst chooses any differencing needed to make the series sufficiently stable, selects terms that use earlier values and earlier forecast errors, adds seasonal terms where a repeatable cycle is supported, estimates the model and checks remaining errors and forecast performance at realistic historical forecast dates. SARIMA also represents a stated seasonal period. This method is the procedure for specifying, fitting and testing one model. The ARIMA / SARIMA framework describes the reusable model family, and neither layer explains the cause of the forecast pattern by itself.
- ARIMAX: Fit an autoregressive integrated moving average model with external inputs (ARIMAX), such as price, promotions, holidays or an economic indicator. Define the forecast target, vintage and horizon, include only predictors that would be available when the forecast is made, use ARIMA errors to represent remaining dependence over time and test performance at later periods. Future predictor values must be known, planned, assumed or separately forecast, and their uncertainty should be carried into the result where material. ARIMAX is one specific form of dynamic regression; the wider Regression / Dynamic Regression method also includes other lag and error structures. A predictive association is not a causal effect unless a separate design supports that claim.
- Backtesting: Recreate how a forecasting or decision procedure would have operated at past decision dates using only information genuinely available then. Define forecast origins, horizons, data revisions, retraining rules, benchmarks and scoring before viewing later outcomes. Backtesting estimates historical prospective performance; leakage from future information or repeated tuning can make it misleading.
- Demand forecasting: Estimate how much of a precisely defined product or service customers are expected to seek in future periods. Specify the unit, customer group, geography, horizon and whether the target is unconstrained demand, orders, sales or served volume; use historical patterns, drivers and primary evidence as available; and report scenarios or prediction uncertainty. Market sizing estimates the total relevant opportunity within a boundary, while a demand forecast describes its expected path through time. Sales may be lower than demand because of share, access, price, capacity or supply constraints.
- Diffusion / adoption modelling: Estimate how members of a defined eligible population begin using an offering over time. Define the adoption event, starting population, time step and saturation level, and represent relevant influences such as awareness, trial, outside promotion, peer effects, access and loss of users. Estimate or justify parameters from observed data, pilots or comparable launches and test sensitivity where evidence is weak. An adoption model describes the timing of first or continuing use under those assumptions. It does not establish the total market boundary, repeat-purchase volume or the sales one provider will capture unless those elements are modelled separately.
- ETS / exponential smoothing: Fit and update an exponential-smoothing forecast in which estimates of the current level, trend and seasonal pattern, where present, respond progressively to new observations. Define the forecast target, seasonal period, vintage and horizon, compare suitable error, trend and seasonal forms, estimate the smoothing parameters and evaluate later-period forecasts against simple benchmarks. Recent observations usually receive more influence than older ones, but the fitted structure determines exactly how information is carried forward. This method applies and tests a particular ETS specification; the ETS / State-Space Forecasting framework groups the reusable model structures and their state-space representation.
- Forecast-combination / ensemble methods: Combine two or more forecasts for the same target, units, forecast vintage and horizon into one point forecast or predictive distribution. Use a stated rule, such as a simple average or weights estimated from prior out-of-sample performance, and prevent actual outcomes from the evaluation period from influencing the weights. Compare the combined forecast with every material constituent and with simple benchmarks across relevant horizons and segments. Combination can reduce errors caused by relying on one model, but it can also repeat common biases or hide a weak component. It combines alternative predictions; it does not turn assumption-based scenarios into probabilities.
- Hierarchical / grouped time-series reconciliation: Adjust base forecasts for related series so that they satisfy defined aggregation or grouping constraints at the same forecast vintage and horizon. Start with a documented hierarchy or grouped structure, such as items within categories and regions within a company, generate or receive the base forecasts, choose a reconciliation rule and use forecast-error information where the rule requires it. Test the coherent results against unreconciled and simple alternatives and show how adjustments affect different levels and uncertainty. The Hierarchical Forecasting framework defines the structure and available reconciliation logic; this method performs the adjustment, and a Hierarchically Reconciled Forecast is the populated result.
- Intermittent-demand methods such as Croston variants: Forecast demand for items or services with many zero-demand periods and irregular non-zero amounts by treating the timing of demand and its size separately or through another method designed for sparse occurrence. Define the item population, time interval, forecast horizon, treatment of stockouts and obsolescence and the loss or service measure used to judge performance. Compare Croston-style variants and other suitable intermittent-demand methods with simple zero, mean and aggregate benchmarks using later forecast periods. The method estimates demand rates or distributions under sparse history; it does not mean that every zero is true absence of demand or that a slow-moving item will continue to be required.
- Interrupted time-series analysis: Estimate whether an outcome’s level or trend changed after an intervention using repeated observations before and after the intervention. Model the pre-intervention trend, immediate level change, post-intervention trend and relevant autocorrelation or seasonality. A change at the intervention point is not necessarily caused by the intervention if other events occurred at the same time; a credible comparison series strengthens attribution.
- Machine-learning forecasting: Use a supervised learning workflow to predict a defined future numerical target from patterns in historical outcomes and other information available at each forecast date. Specify the target, units, population, vintage and horizon; create lag, calendar and known-future features without leakage; choose how multiple horizons are produced; train and tune models on earlier periods; and evaluate them on later periods against simple and established forecasting methods. This is an umbrella method that can include gradient boosting, random forests, neural networks and other algorithms. It produces predictive relationships rather than causal effects, and additional model complexity is justified only by useful out-of-sample performance and acceptable operating controls.
- Market-potential demand forecasting: Forecast how demand for a market, category or proposition could develop over time, from its size, adoption path and drivers, typically before an organisation has its own sales history. It is a specialisation of demand forecasting aimed at market potential. It estimates what the market could support, not the dated operational sales an existing business should plan for.
Parent method family
Forecasting, Risk & Optimisation explains how this method connects to adjacent methods and relevant services.
Related service families
These service families contain business questions supported by this method. Service pages link to the wider method family so readers can understand the complete analytical approach.