Time-Series Forecasting Models

Introduction and context

Supply-chain decisions are made before the future is known. A retailer must decide how much inventory to position, a manufacturer must schedule capacity and materials, and a distributor must plan transport and labour. Time-series forecasting supports these decisions by using the historical behaviour of a variable—usually demand, sales, orders, lead time, or inventory transactions—to estimate its future values.

A time series is an ordered sequence of observations indexed by time. Let y_t denote the observed value in period t, where the period may be a day, week, month, or quarter. Forecasting is the process of estimating y_t+h for a future horizon h.

Time-series models are especially useful when the history of the variable contains information about its future. They are not substitutes for managerial knowledge: promotions, product launches, competitor actions, supply disruptions, and structural changes may not be visible in the historical series. Effective supply-chain forecasting therefore combines statistical models, causal information, judgement, and disciplined performance monitoring.

Learning objectives

After studying this chapter, the learner should be able to:

1. Explain level, trend, seasonality, cycle, noise, autocorrelation, and stationarity.
2. Develop naïve, moving-average, weighted-average, exponential-smoothing, Holt, Holt–Winters, decomposition, and ARIMA forecasts.
3. Select smoothing parameters and model orders using appropriate evidence.
4. Calculate and interpret forecast errors, prediction intervals, and bias.
5. Implement and evaluate models using spreadsheets or analytical software.
6. Recognise model assumptions, limitations, governance risks, and supply-chain implications.

1. Structure of a demand time series

A useful conceptual representation is

where L_t is the underlying level, T_t is trend, S_t is seasonal movement, C_t is a longer cycle, and ε_t is irregular error. This additive form is appropriate when seasonal variation is approximately constant in absolute units. A multiplicative representation is

and is appropriate when seasonal amplitude increases with the level. Taking logarithms converts a multiplicative relationship into an additive one, provided observations are positive:

The principal components are:

• Level: the current typical value.
• Trend: persistent upward or downward movement.
• Seasonality: a regular pattern repeating at a known frequency, such as weekly or annual demand.
• Cycle: a fluctuation with a less fixed duration, often associated with economic conditions.
• Noise: unpredictable variation.
• Outlier: an observation unusually distant from the prevailing pattern.
• Autocorrelation: dependence between observations separated by a lag.
• Stationarity: a process whose probabilistic properties, including mean and variance, do not systematically change over time.

Forecasting must distinguish calendar seasonality from business events. A weekly pattern is not necessarily the same as a quarterly pattern, and a one-time stockout should not automatically be treated as low customer demand.

2. Forecasting workflow

A robust workflow has six stages:

1. Define the decision and horizon. Forecasting a replenishment quantity for the next week differs from forecasting annual capacity.
2. Prepare the data. Establish a consistent time index, aggregate to the relevant frequency, identify missing periods, correct duplicate records, and document stockouts and promotions.
3. Explore the series. Plot the history, seasonal subseries, rolling means, rolling standard deviations, and autocorrelation.
4. Create benchmark forecasts. Naïve and seasonal-naïve models provide essential reference points.
5. Estimate and validate candidate models. Use time-ordered training and validation periods rather than random cross-validation.
6. Deploy and monitor. Track errors, bias, parameter stability, forecast overrides, and changes in the business environment.

Data leakage is a major risk. Information unavailable at the forecast origin must not enter model estimation or validation. For example, using the final month’s demand to select a model for that same month produces an unrealistically favourable result.

3. Baseline models

3.1 Naïve forecasting

The simplest forecast assumes that the next value equals the most recent observation:

Here F_t+1 is the forecast made at time t for period t+1. The method is appropriate as a benchmark and can be effective for short-horizon series whose level changes slowly. It does not explicitly model trend or seasonality.

A seasonal-naïve forecast repeats the observation from the equivalent season:

where m is the seasonal period, such as m=12 for monthly annual seasonality or m=7 for daily weekly seasonality. Seasonal naïve forecasting is often a demanding benchmark for seasonal supply-chain demand.

3.2 Simple moving average

A moving average of order k forecasts the next period using the mean of the most recent k observations:

All observations receive equal weight. A larger k provides greater smoothing but responds more slowly to genuine changes. A smaller k reacts faster but retains more noise.

Example. If demand in the last four weeks is 96, 104, 100, and 108 units, the four-week moving-average forecast is

The forecast is in units because it averages quantities measured in units.

3.3 Weighted moving average

A weighted moving average assigns greater importance to selected observations:

where each weight is non-negative and the weights sum to one:

If the last three demands are 120, 110, and 100, with weights 0.5, 0.3, and 0.2 respectively, then

Weights may be chosen by managerial judgement or estimated by minimising a validation loss. They should not be selected merely because they fit the historical sample particularly well.

4. Exponential smoothing

4.1 Simple exponential smoothing

Simple exponential smoothing (SES) is designed for series with no systematic trend or seasonality. It updates the level recursively:

The forecast for all future periods is

where ℓ_t is the estimated level and α is the smoothing parameter. A high α reacts rapidly to new observations; a low value produces stronger smoothing. The initial level ℓ₀ must be specified, often as the first observation or an initial average.

SES can also be written as a weighted average of past observations, with weights declining geometrically. Thus it uses all available history while emphasising recent information.

4.2 Parameter estimation

A parameter may be selected by minimising a loss function over a training sample. A common objective is the sum of squared one-step-ahead errors:

where n is the number of evaluated observations. Mean squared error is MSE=SSE÷n. Optimising squared error penalises large errors strongly; mean absolute error may be preferable when robustness and interpretability are more important.

Parameter optimisation should be performed within the training data. A separate validation period is used to assess generalisation. Software may optimise both smoothing parameters and initial states, but the analyst remains responsible for checking plausibility and stability.

5. Trend models: Holt’s method

Simple exponential smoothing cannot represent a continuing trend adequately. Holt’s linear method estimates both level and trend:

Here b_t is the estimated trend and β satisfies 0<β<1. The first equation compares the current observation with the previous level extrapolated by the previous trend. The second updates the trend. Forecasts are linear in the horizon and can become unrealistic for long horizons, especially for bounded or intermittent demand.

Damped trend

A damped trend reduces the extrapolated trend as the horizon increases:

where 0<φ<1. Damped trends are often useful in supply chains because indefinite exponential or linear growth is rarely credible. The value of φ may be estimated from historical performance.

6. Seasonality: Holt–Winters methods

Holt–Winters methods extend exponential smoothing to level, trend, and seasonality. Let m denote the number of observations per seasonal cycle.

6.1 Additive seasonality

For approximately constant seasonal amplitude:

where 0<α,β,γ<1, s_t is the seasonal component, and h^* maps the forecast horizon to the relevant seasonal position. In implementation, the final term is indexed cyclically through the seasonal vector.

6.2 Multiplicative seasonality

For seasonal amplitude proportional to the level:

Multiplicative models require positive observations and are inappropriate when zeros or negative values occur unless a carefully justified transformation is used. Additive models are safer when demand variation is measured in roughly constant units.

7. Classical decomposition

Decomposition separates a series into interpretable components. In additive decomposition:

In multiplicative decomposition:

where T_t is trend-cycle, S_t is seasonality, and R_t is the remainder. A common procedure estimates trend using a centred moving average, estimates seasonal indices from detrended observations, and forecasts the trend before recombining it with the seasonal pattern.

For multiplicative seasonality, seasonal indices are commonly normalised so that their average over a full cycle is one. For additive seasonality, indices are typically constrained to sum to zero over a full cycle. These constraints prevent the level from being counted twice.

Decomposition is valuable for explanation and diagnostics, but classical decomposition is less adaptive than state-space exponential smoothing when patterns evolve. Modern decomposition methods may use robust trend extraction, moving seasonal windows, or transformations.

8. ARIMA models

8.1 Autoregressive and moving-average terms

An autoregressive model of order p uses prior observations or transformed observations:

An MA model of order q uses current and prior shocks:

The innovations ε_t are commonly assumed to have mean zero, constant variance, and no serial correlation. They need not be normally distributed for point forecasts to be useful, although normality is often used for simple prediction intervals.

8.2 ARIMA(p,d,q)

ARIMA combines autoregression, differencing, and moving-average terms. If B is the backshift operator, defined by Bx_t=x_t−1, an ARIMA model may be written as

where p is the AR order, d is the number of differences, q is the MA order, and φ(B) and θ(B) are polynomial operators. Differencing is used to remove stochastic trend and achieve approximate stationarity. Over-differencing can create unnecessary noise and distort the dependence structure.

If seasonality is present, seasonal ARIMA extends the model with seasonal orders (P,D,Q)_m, where m is the seasonal period. Model identification may use plots, autocorrelation and partial autocorrelation functions, information criteria, and operational knowledge. AIC and corrected AIC balance fit against parameter count, but they do not replace out-of-sample evaluation.

8.3 Residual diagnostics

After fitting an ARIMA model, residuals should approximate white noise. Important checks include:

• residual mean close to zero;
• no material autocorrelation;
• stable variance;
• absence of unexplained seasonal structure;
• plausible out-of-sample accuracy.

The Ljung–Box test can assess whether a group of residual autocorrelations is jointly different from zero. A non-significant result does not prove that the model is correct; it only indicates insufficient evidence of remaining serial correlation under the test assumptions.

9. Forecast accuracy and prediction intervals

Let e_t=y_t−F_t be the forecast error. Common measures include:

MAE is in the original units and is readily interpreted. RMSE penalises large errors more heavily. MAPE is scale-free but is undefined when actual demand is zero and unstable when actual demand is close to zero. Weighted absolute percentage error or mean absolute scaled error can be preferable for intermittent or low-volume items.

Forecast bias can be monitored using mean error:

A positive mean error under the stated convention means forecasts are systematically below actual demand. Bias matters operationally: persistent underforecasting can cause stockouts, while persistent overforecasting can create excess inventory.

A point forecast is not a guarantee. A prediction interval describes uncertainty about a future observation under model assumptions. A simple approximate interval is

where σ_h is the horizon-specific forecast-error standard deviation, α is the probability outside the interval, and z_1−α÷2 is a standard-normal quantile. For example, a nominal 95% interval uses approximately z=1.96. In many time-series models, σ_h increases with horizon and is not equal to the residual standard deviation for every h. Intervals may also be asymmetric when demand is non-negative, intermittent, or model uncertainty is substantial.

Prediction intervals should not be confused with confidence intervals for an estimated mean. The former concern a future observation; the latter concern an unknown parameter or expected value.

10. Model selection and implementation

A practical model-selection hierarchy is:

1. establish naïve and seasonal-naïve benchmarks;
2. inspect trend and seasonality;
3. fit SES, Holt, or Holt–Winters where appropriate;
4. consider ARIMA when autocorrelation and differencing are important;
5. compare models using rolling-origin validation;
6. examine residuals, bias, stability, and business plausibility;
7. select or combine models rather than relying solely on the lowest in-sample error.

Rolling-origin validation repeatedly trains on an expanding or moving historical window and forecasts the next period or several periods. This reflects the actual forecasting task and is generally more credible than random sampling for ordered data.

In a spreadsheet, each row should represent a time period and columns should contain actual demand, forecast, error, absolute error, squared error, and any explanatory flags. In statistical software, the analyst should document frequency, missing-value treatment, transformations, initialisation, optimisation method, training cut-off, validation design, and software version.

Forecast combinations can improve robustness. For example, a weighted combination of a seasonal-naïve forecast and an exponential-smoothing forecast is

The combination weight may be equal across models or estimated using validation data. Combining does not eliminate the need for monitoring; correlated model errors may limit its benefit.

11. Illustrative case: Northstar Home Appliances

Background

Northstar Home Appliances sells a replacement water-filter cartridge through retailers and an online channel. Weekly demand over three years shows a mild upward trend, a recurring increase before holidays, and occasional zero sales at small retailers. The planning team currently uses the previous week’s demand for replenishment.

Problem

The previous-week method reacts strongly to promotions and stockouts. It also ignores the annual holiday pattern. In the last validation quarter, the method produced a MAE of 84 units and a mean error of −46 units, indicating systematic underforecasting under the convention e_t=y_t−F_t. The resulting service failures required emergency shipments.

Analysis

The team first separated retailer stockouts from genuine demand. It aggregated online and retail demand to a weekly total and created a 52-week seasonal index. Candidate models were seasonal naïve, SES, Holt’s damped trend, additive Holt–Winters, and a seasonal ARIMA model. The final 13 weeks were held out for validation, and earlier observations were used for estimation. Model comparison considered MAE, RMSE, bias, interval coverage, and operational interpretability.

Approach and solution

The seasonal-naïve model captured the holiday pattern but missed the recent growth. SES responded to growth but did not represent seasonality. Additive Holt–Winters produced the best balance: its seasonal amplitude was approximately stable in units, residual autocorrelation was small, and its validation bias was close to zero. The planning team therefore selected it as the primary model and retained seasonal naïve as a monitoring benchmark.

The forecast was converted into a replenishment requirement using inventory policy. If lead-time demand forecast is μ_L and the desired safety factor is z, a simplified safety-stock expression is

where σ_L is the standard deviation of lead-time demand. The reorder point was then set as

These inventory equations assume that lead-time demand uncertainty is represented appropriately and that the selected service factor corresponds to the organisation’s service objective. They are not a substitute for modelling demand–lead-time dependence or intermittent demand.

Outcome and learning points

After two review cycles, Northstar observed lower bias and fewer emergency shipments. The improvement was not attributed to the model alone: correcting stockout data, using a seasonal benchmark, and establishing a weekly exception review were equally important. The case illustrates five principles: forecast the true demand signal, validate chronologically, assess bias as well as accuracy, connect forecast uncertainty to inventory policy, and retain human review for promotions and structural changes.

12. Limitations, governance, and contemporary practice

Time-series models assume that historical relationships provide useful information about the future. Structural breaks, new products, cannibalisation, market shocks, policy changes, and supply constraints can invalidate this assumption. Observed sales may be censored by stockouts; forecasting sales without correcting for lost demand can systematically understate market demand.

Governance requires traceability. Organisations should record data provenance, transformations, model versions, overrides, approval responsibilities, and performance by product and location. Forecast overrides should include a reason code and later be evaluated; otherwise judgement can be praised or blamed without evidence.

Machine-learning systems can incorporate promotions, price, weather, events, web traffic, and hierarchical relationships. They are useful when external drivers contain stable predictive information and sufficient history exists. However, a more complex algorithm is not automatically superior to a well-calibrated benchmark. Explainability, drift monitoring, data access controls, and fairness across customer or geographic segments are essential.

Best practice is to use a model portfolio: simple benchmarks for control, statistical models for repeatable baseline forecasts, causal or machine-learning models where drivers justify them, and structured judgement for exceptional information. Forecasts should be evaluated at the level at which decisions are made, while reconciliation may be required so that forecasts for products, locations, channels, and total demand are coherent.

Chapter Summary

Time-series forecasting estimates future observations from ordered historical data. Demand may contain level, trend, seasonality, cycles, and irregular noise. Naïve and seasonal-naïve forecasts establish essential benchmarks. Moving averages smooth recent observations, while weighted averages emphasise selected periods. Simple exponential smoothing estimates a changing level; Holt’s method adds trend; damped trend limits unrealistic extrapolation; and Holt–Winters adds additive or multiplicative seasonality. Classical decomposition separates trend, seasonal, and remainder components for interpretation. ARIMA models represent dependence through autoregression, differencing, and moving-average errors, with seasonal extensions for recurring patterns.

Model parameters must be estimated without leakage and evaluated using time-ordered validation. MAE, RMSE, percentage measures, and bias answer different managerial questions. Prediction intervals quantify uncertainty and generally widen with the forecast horizon. Residual diagnostics, operational plausibility, data quality, and monitoring are as important as numerical fit. In supply chains, forecasts inform replenishment and capacity but must be integrated with lead times, service objectives, promotions, stockouts, governance, and expert judgement.

Key Terms

Autocorrelation: Dependence between observations at different time lags.

ARIMA: A model combining autoregression, differencing, and moving-average error terms.

Bias: Systematic tendency for forecasts to be above or below actual values.

Decomposition: Separation of a series into components such as trend, seasonality, and remainder.

Damped trend: A trend forecast whose extrapolation declines with the horizon.

Exponential smoothing: Recursive forecasting that assigns geometrically declining weights to past observations.

Forecast origin: The time at which a forecast is produced.

Holt’s method: Exponential smoothing for level and trend.

Holt–Winters method: Exponential smoothing for level, trend, and seasonality.

Innovation: A new, unpredictable error or shock in a time-series model.

Moving average: A forecast based on the mean of recent observations.

Naïve forecast: A forecast that repeats the latest observation.

Prediction interval: A range intended to contain a future observation with stated probability.

Seasonality: A pattern that repeats at a known calendar frequency.

Seasonal naïve forecast: A forecast that repeats the observation from the equivalent prior season.

Stationarity: Stability of a process’s probabilistic properties over time.

Time series: Observations ordered by time.

Validation: Evaluation of a model on data not used for its estimation.

Review Questions

Short Answer Questions

1. Distinguish trend from seasonality in a supply-chain demand series.
2. Why is a seasonal-naïve forecast an important benchmark?
3. What does the smoothing parameter α control in simple exponential smoothing?
4. State one advantage and one limitation of MAPE.
5. Why should residuals from an ARIMA model be approximately uncorrelated?

Long Answer Questions

1. Explain the assumptions, equations, strengths, and limitations of simple exponential smoothing, Holt’s method, and Holt–Winters methods.
2. Compare classical decomposition with ARIMA forecasting for a seasonal product.
3. Explain how rolling-origin validation should be designed for a monthly supply-chain forecast.
4. Discuss why forecast bias may be more operationally important than a small difference in RMSE.
5. Describe a governance framework for deploying statistical forecasts across products and locations.

Analytical / Application Questions

1. Weekly demand for the last four weeks is 96, 104, 100, and 108 units. Calculate the four-week moving-average forecast. If weights 0.5, 0.3, and 0.2 are assigned to the most recent three weeks, calculate the weighted forecast.
2. A product has monthly demand with stable seasonal amplitude and a modest trend. Recommend a model from those covered in the chapter, justify the choice, and specify two diagnostics you would perform before deployment.
3. A forecast system has MAE of 40 units but mean error of −28 units. Interpret the bias, explain two possible supply-chain consequences, and propose a monitoring or corrective action.

Multiple Choice Questions with Answers

Question 1

Which model is most appropriate as a basic benchmark for demand with a stable annual seasonal pattern?

A. Seasonal-naïve forecast

B. Unrestricted linear trend

C. Randomly shuffled average

D. Zero-demand forecast

Correct Answer: A. Seasonal-naïve forecast

Question 2

A four-period moving average uses demands 90, 100, 110, and 120 units. What is the next forecast?

A. 100 units

B. 105 units

C. 110 units

D. 115 units

Correct Answer: B. 105 units

Question 3

In simple exponential smoothing, a larger α generally makes the forecast:

A. Less responsive to recent observations

B. More responsive to recent observations

C. Independent of the latest observation

D. Automatically seasonal

Correct Answer: B. More responsive to recent observations

Question 4

Which method explicitly estimates both a level and a trend?

A. Simple exponential smoothing

B. Holt’s linear method

C. Seasonal naïve forecasting

D. Unweighted moving average only

Correct Answer: B. Holt’s linear method

Question 5

Multiplicative Holt–Winters seasonality is most appropriate when seasonal variation:

A. Is approximately constant in absolute units

B. Is proportional to the level of the series

C. Occurs only once

D. Contains negative observations by definition

Correct Answer: B. Is proportional to the level of the series

Question 6

Under the error definition e_t=y_t−F_t, a negative mean error indicates that forecasts are, on average:

A. Below actual demand

B. Equal to actual demand

C. Above actual demand

D. Always zero

Correct Answer: C. Above actual demand

Question 7

What is the main purpose of differencing in ARIMA modelling?

A. To introduce deterministic seasonality

B. To help remove stochastic trend and achieve stationarity

C. To guarantee zero forecast error

D. To replace all residual diagnostics

Correct Answer: B. To help remove stochastic trend and achieve stationarity

Question 8

Why is random train-test splitting generally unsuitable for ordinary time-series validation?

A. It prevents calculation of MAE

B. It can use future information to predict the past

C. It always increases sample size

D. It removes seasonality automatically

Correct Answer: B. It can use future information to predict the past

Question 9

A 95% normal-based approximate prediction interval commonly uses which multiplier?

A. 0.50

B. 1.00

C. 1.645

D. 1.96

Correct Answer: D. 1.96

Question 10

If demand is repeatedly censored by stockouts, using observed sales without adjustment may:

A. Systematically overstate unconstrained demand

B. Systematically understate unconstrained demand

C. Remove all forecast bias

D. Make seasonality impossible

Correct Answer: B. Systematically understate unconstrained demand