OperationsFramework / modelModelAccessible

Forecasting/time series analysis

How can forecasting/time series analysis improve people, teams, or organisational effectiveness?

AccessibleOperationalOrganisation2 min read

Contents

To understand what time series analysis is you must first understand what time series data is.

A time series is a sequence of observations indexed in time. Time-series analysis describes how that sequence changes, separates recurring structure from irregular variation and, where appropriate, forecasts future values with quantified uncertainty.

When to use it

Use time-series analysis when the order and spacing of observations matter—for example, daily demand, monthly revenue, hourly energy use or a market closing value. The method can describe trend, seasonality, cycles, abrupt changes and dependence between current and earlier observations.

It can support questions such as:

How might economic or business performance evolve over the coming months?

What production volume and inventory may be needed?

Is a recent movement normal seasonal variation or a structural break?

How uncertain is the forecast, and which decisions are sensitive to that uncertainty?

Use another design when there is too little history, the process has changed fundamentally or the main objective is to estimate the causal effect of an intervention.

Origins

Time-series methods developed across astronomy, economics, signal processing and statistics. Early researchers studied trends, periodicity and serial correlation; later work formalised autoregressive and moving-average processes. George Box and Gwilym Jenkins integrated model identification, estimation, diagnostic checking and forecasting into an influential practical workflow. Contemporary analysis includes exponential smoothing, state-space models, dynamic regression and machine-learning approaches, but the same discipline remains: respect temporal order and validate forecasts out of sample.

What it is

Time-series analysis treats observations as potentially dependent rather than interchangeable. A value today may be related to recent values, seasonal positions, external variables and shocks.

A useful decomposition distinguishes:

level: the current baseline of the series;

trend: sustained movement over time;

seasonality: a pattern repeating at a known calendar or operational frequency;

cycle: broader movement without a fixed seasonal period;

irregular variation: shocks, noise and measurement error.

Forecasts extrapolate learned structure under explicit assumptions. They are conditional estimates, not predictions that the future will repeat the past exactly.

How to use it

Define the decision, forecast horizon, update cadence and error cost. A staffing forecast and a capital plan may require different horizons and loss functions.

Assemble time-stamped data and check frequency, missing observations, revisions, outliers and changes in definition. Do not automatically delete unusual values: correct genuine errors, but retain real events and decide whether the future could contain comparable shocks.

Plot the series and compare it with relevant calendars and external variables. Create a simple benchmark such as the last observed value or a seasonal-naive forecast before fitting a more complex model.

Candidate approaches include autoregressive models, moving-average models, their integrated combinations, exponential smoothing and dynamic regression. Fit models only on past data, evaluate them on later holdout periods or rolling forecast origins, inspect residuals and report prediction intervals as well as point forecasts.

Monitor error after deployment. Refit or redesign when drift, new competition, regulation, product changes or another structural break makes earlier relationships unreliable.

Practical example

Suppose a company needs quarterly product-sales forecasts to plan production, materials, inventory, warehousing and targets. Historical sales show an upward trend and a recurring seasonal peak.

The analyst compares a seasonal-naive benchmark with candidate models, reserves the latest periods for validation and tests whether promotions, price or availability improve accuracy. The final forecast includes a range and scenarios for an unusual market change. Operations can then plan a central case while defining actions for higher or lower demand.

Top practical tip

Always compare against a simple benchmark on genuinely later data. Track error by horizon and segment, show prediction intervals and keep a decision log. A sophisticated model adds value only if it improves the forecast or the decision after complexity and maintenance are considered.

Top pitfall

Do not confuse recurring history with a stable future. Leakage from future data, revised definitions, one-off shocks and structural change can make validation look better than deployment. Time-series association also does not establish cause. When the market is changing, combine statistical forecasts with explicit scenarios and update frequently.