Time-Series Forecasting Models

Introduction and Context

Time-series forecasting estimates future values from observations ordered through time. In supply chain management, forecasts support procurement, production, inventory positioning, transportation capacity, workforce planning, budgeting, and customer-service commitments. A time series may contain several distinguishable features:

• Level: the typical magnitude around which observations fluctuate.
• Trend: persistent upward or downward movement.
• Seasonality: a repeating pattern with a known or stable period, such as weekly, monthly, or annual demand cycles.
• Cycle: longer, often less regular expansion and contraction associated with economic or industry conditions.
• Irregular variation: short-term noise, shocks, and measurement error.

The central forecasting challenge is not simply to fit historical data. It is to identify which components are sufficiently stable to extrapolate, while avoiding overreaction to random variation and structural breaks.

Learning Objectives

By the end of this chapter, students should be able to:

1. Explain the assumptions and appropriate uses of major time-series forecasting models.
2. Calculate and interpret naïve forecasts, moving averages, exponential smoothing, Holt trend, and Holt–Winters forecasts.
3. Distinguish additive and multiplicative decomposition and relate model choice to demand behavior.
4. Compare models using out-of-sample error measures and rolling-origin evaluation.
5. Describe the logic, notation, assumptions, and limitations of introductory ARIMA models.
6. Select and govern forecasting methods in practical supply chain settings.

1. Forecasting as a Time-Ordered Inference Problem

Let [[M:y_t]] denote observed demand in period [[M:t]], where [[M:t = 1, 2, …, n]]. A forecast made at time [[M:t]] for [[M:h]] periods ahead is written [[M:\hat{y}_{t+h|t}]]. The notation emphasizes that the forecast uses information available at time [[M:t]], not future observations.

A forecast is therefore a conditional estimate, not a guaranteed outcome. In operational planning, the forecast should be distinguished from the plan: a forecast describes likely demand, whereas a plan may deliberately constrain, stimulate, or substitute for demand.

A useful decomposition is:

for an additive representation, where [[M:T_t]] is trend, [[M:S_t]] is seasonal effect, and [[M:R_t]] is the remainder. A multiplicative representation is:

The additive form is appropriate when seasonal fluctuations have approximately constant absolute size. The multiplicative form is appropriate when seasonal variation grows with the level. Multiplicative data must generally be positive; a logarithmic transformation can convert multiplication into addition when its assumptions are reasonable.

2. Establishing a Sound Forecasting Workflow

A professional forecasting process normally includes the following steps:

1. Define the forecasting decision. Specify the item, location, time bucket, horizon, decision owner, and required forecast frequency.
2. Audit the data. Check missing periods, duplicate records, stockouts, returns, promotions, price changes, calendar effects, and changes in product coding.
3. Visualize the series. Plot demand over time and inspect level, trend, seasonality, outliers, and variance changes.
4. Create benchmark forecasts. Naïve methods provide difficult-to-beat baselines and reveal whether complex modeling adds value.
5. Fit candidate models. Use methods consistent with the observed pattern and business context.
6. Validate chronologically. Train on earlier observations and evaluate on later observations; random shuffling usually leaks future information.
7. Measure uncertainty. Produce prediction intervals where possible and communicate assumptions.
8. Monitor and revise. Track forecast error, bias, parameter stability, and process changes.

A forecast accuracy study should use a holdout period or rolling-origin evaluation. A model that fits historical data exceptionally well may perform poorly when the environment changes.

3. Baseline and Smoothing Models

3.1 The naïve model

The simplest forecast for a nonseasonal series is the last-observation forecast:

For a seasonal naïve model with seasonal period [[M:m]], the forecast repeats the observation from the same season in the previous cycle:

when the required historical value exists. The seasonal naïve method is a strong benchmark for stable seasonal demand. It assumes that the most recent comparable season is more informative than older seasons and that seasonal structure persists.

Example. Weekly demand for the last four corresponding weeks was 94, 101, 98, and 105 units. A naïve forecast for the next week is 105 units. If the series has a four-week cycle and the corresponding observation one cycle earlier was 98, the seasonal naïve forecast is 98 units instead.

The naïve method is transparent, inexpensive, and often effective for short horizons. It does not explicitly estimate trend, seasonality, or noise, and it can be unstable after an unusual one-period shock.

3.2 Moving averages

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

Here, [[M:k]] is a positive integer and all observations must use the same units. The method assumes that recent observations are representative of the near future and gives equal weight to the selected periods.

Worked example. Suppose the most recent four monthly demands are 120, 135, 150, and 145 cases. A four-month moving-average forecast is:

A larger [[M:k]] produces greater smoothing and less responsiveness. A smaller [[M:k]] responds more quickly but retains more random variation. A moving average lags a changing trend because historical low values remain in the window after demand begins to rise. It also does not naturally account for seasonality unless the window and treatment are specifically designed for it.

3.3 Weighted moving averages

A weighted moving average assigns unequal weights to recent observations:

The weights are dimensionless and the forecast retains the units of [[M:y]]. Greater weight on recent observations improves responsiveness but may increase sensitivity to noise. Weight selection should be validated rather than based only on intuition.

4. Decomposition and Seasonal Structure

Decomposition separates a series into interpretable components. It is useful diagnostically and can also generate forecasts by extrapolating the estimated trend and repeating seasonal factors.

4.1 Additive decomposition

An additive model is suitable when the seasonal amplitude is approximately constant. For example, a product may sell about 200 additional units every December regardless of its baseline level.

4.2 Multiplicative decomposition

A multiplicative model is suitable when seasonal amplitude is proportional to the level. For example, December demand might consistently be 1.40 times average demand. Seasonal indices are commonly normalized so that their mean over a full cycle equals 1.00. In an additive model, seasonal effects over a complete cycle commonly sum to zero.

A moving-average decomposition estimates the trend-cycle component by smoothing observations, estimates seasonal effects from deviations relative to that component, and assigns remaining variation to the residual. More advanced decompositions, such as STL, permit robust handling of outliers and changing seasonal patterns.

Interpretive example. If baseline monthly demand rises from 1,000 to 1,500 units and the December seasonal index is 1.30, a multiplicative estimate for December before adding residual effects is:

The result is meaningful only if the index was estimated from comparable data and the trend and seasonal relationship remains stable.

Decomposition does not prove that a causal seasonal mechanism exists. A calendar pattern may reflect promotions, operating schedules, billing cycles, or omitted variables. Managers should investigate causes before using the pattern for strategic commitments.

5. Simple Exponential Smoothing

Simple exponential smoothing (SES) is designed for series with a changing level but no systematic trend or seasonality. The forecast is updated recursively:

and future forecasts are:

where [[M:\ell_t]] is the estimated level, [[M:0 < \alpha < 1]] is the smoothing parameter, and [[M:h]] is the forecast horizon. All values have the same demand units; [[M:\alpha]] is unitless.

The equivalent error-correction form is:

Thus, the level moves in the direction of the latest forecast error. A high [[M:\alpha]] reacts strongly to new information; a low [[M:\alpha]] produces a stable but slower-moving estimate.

Worked example. Let the previous level be 100 units, current demand be 112 units, and [[M:\alpha = 0.30]]. Then:

The one- through multi-period forecasts are 103.6 units under the basic SES model. Initialization matters, especially in short series. Common choices include setting the initial level to the first observation, an early average, or an estimated value; the choice should be documented and tested.

SES is not a trend model. Applying it to a strongly trending series creates systematic lag and bias.

6. Holt’s Linear Trend Method

Holt’s method extends exponential smoothing with a level and trend state:

Here [[M:b_t]] is the estimated change per period, [[M:0 < \beta < 1]] is the trend-smoothing parameter, and [[M:\alpha]] is the level-smoothing parameter. The forecast has demand units, while [[M:b_t]] has demand units per period.

Higher [[M:\beta]] allows the estimated trend to change rapidly. Lower [[M:\beta]] stabilizes the trend. The un-damped method extrapolates the trend indefinitely and can produce unrealistic forecasts over long horizons. A damped trend introduces a factor [[M:0 < \phi < 1]]:

As [[M:h]] increases, the trend contribution grows more slowly than under a linear extrapolation. Damping is often sensible when current growth is unlikely to continue permanently.

7. Holt–Winters Seasonal Models

Holt–Winters methods combine level, trend, and seasonality. Let [[M:m]] be the number of observations per seasonal cycle.

7.1 Additive Holt–Winters

For approximately constant seasonal amplitude:

The final expression cycles through the appropriate seasonal states. Parameters [[M:\alpha]], [[M:\beta]], and [[M:\gamma]] lie between 0 and 1. The seasonal component uses the same units as demand in the additive model.

7.2 Multiplicative Holt–Winters

For seasonality proportional to the level:

Multiplicative seasonality requires positive, suitably scaled observations. The seasonal factor is dimensionless. If demand contains zeros, a multiplicative formulation requires careful transformation or an alternative model; replacing zeros casually can distort results.

Parameter estimation may minimize a training loss such as the sum of squared one-step errors, but model selection must still rely on temporal validation. Complex seasonal models can overfit when only a few cycles are available.

8. Forecast Error and Model Comparison

Let [[M:e_t = y_t - \hat{y}_t]] be the forecast error. Positive error means actual demand exceeded forecast, so the model underforecast.

Common measures include:

MAE is expressed in demand units and is easy to interpret. RMSE penalizes large errors more heavily. MAPE is scale-free but is undefined when [[M:y_t = 0]] and unstable when actual values are close to zero. Alternatives include weighted absolute percentage error or a scaled error such as MASE:

where the denominator is calculated from an appropriate in-sample benchmark, commonly a naïve or seasonal-naïve difference. The exact benchmark must be stated.

Accuracy is not the only criterion. A supply chain may prefer a slightly less accurate model that is explainable, stable, fast, and compatible with inventory policy. Examine bias, service-level consequences, forecast intervals, and costs of over- versus under-forecasting.

9. Introductory ARIMA Concepts

ARIMA models describe dependence among observations and forecast errors after suitable differencing. The notation ARIMA([[M:p,d,q]]) contains:

• [[M:p]]: autoregressive order;
• [[M:d]]: number of differences used to achieve approximate stationarity;
• [[M:q]]: moving-average order of the error process.

An AR(1) model is:

where [[M:c]] is a constant, [[M:\varphi]] is the autoregressive coefficient, and [[M:\varepsilon_t]] is an unpredictable error with mean approximately zero. For weak stationarity of a basic AR(1), a common condition is [[M:|\varphi| < 1]].

An MA(1) model is:

where [[M:\mu]] is the mean and [[M:\theta]] controls the effect of the previous shock. The “moving average” here refers to past errors, not averages of past observations.

For a nonseasonal first difference, define:

An ARIMA model applies ARMA structure to differenced data. Stationarity means that key statistical properties, such as mean and autocovariance, do not change with time. Differencing can remove stochastic trend, but over-differencing can introduce unnecessary dependence and inflate forecast uncertainty.

Model identification commonly considers the time plot, differenced plot, autocorrelation function (ACF), and partial autocorrelation function (PACF). ACF/PACF patterns are guides, not mechanical proof. Candidate models should be estimated, checked for residual autocorrelation, compared with information criteria such as AICc, and evaluated on withheld future observations.

A seasonal ARIMA model extends the orders to seasonal autoregressive, differencing, and moving-average terms. External drivers such as price, promotions, weather, or economic indicators require an explanatory extension, often called dynamic regression or ARIMAX. Pure ARIMA cannot discover causal information absent from the time series.

10. Selecting Models in Supply Chain Practice

A practical selection logic is:

Forecast combinations can improve robustness when models make different errors. A simple combination is the arithmetic mean of forecasts, although weights may be optimized using validation data. Human overrides should be recorded with reason codes and later audited; otherwise, apparent forecast improvement may reflect undocumented optimism rather than model quality.

11. Illustrative Case: Northstar Components

Background. Northstar Components supplies replacement filters to regional maintenance distributors. Monthly demand over three years shows a gradual increase, a recurring year-end peak, and occasional stockouts. The planning team currently uses a three-month moving average.

Problem. The moving average underforecasts November and December and reacts slowly after a new service contract increases the customer base. However, management does not want an aggressive trend extrapolation that could create excess inventory if the contract is not renewed.

Analysis. The team first removes periods affected by stockouts from the demand analysis or estimates lost demand using documented procedures. A time plot indicates both trend and seasonality. Seasonal amplitude rises with the level, suggesting multiplicative rather than additive seasonality. The team compares seasonal naïve, damped Holt, multiplicative Holt–Winters, and a seasonal ARIMA candidate.

The last 12 months are held out, and earlier data are used for estimation. Models are assessed using MAE, a scaled error, bias, and inventory simulation. The current three-month average has good short-term smoothness but systematically underforecasts peak months. The seasonal naïve model captures the calendar pattern but misses the contract-related level increase. Multiplicative Holt–Winters captures both features, while damped trend without seasonality performs poorly in peak months. The ARIMA candidate has similar average accuracy but less stable results across rolling-origin windows.

Approach and solution. Northstar adopts multiplicative Holt–Winters as the baseline forecast, with a damped trend and a documented adjustment for the service contract. The adjustment is treated as a scenario rather than silently embedded in the statistical model. Planners produce a central forecast and an uncertainty range, then translate the range into inventory and capacity decisions. Forecasts are monitored monthly, with separate tracking for ordinary demand, promotion effects, and stockout-induced bias.

Outcome. The approach reduces seasonal underforecasting and makes assumptions visible. It does not eliminate uncertainty: contract renewal, supplier lead-time variability, and customer substitution remain material risks.

Learning points. The best model was selected through temporal validation and operational consequences, not training fit alone. Data quality and event labeling were as important as the algorithm. A forecast model should coexist with scenario planning when structural changes are plausible.

12. Limitations, Governance, and Contemporary Practice

Time-series models assume some persistence in the data-generating process. They can fail after pandemics, regulatory changes, product launches, supplier exits, channel shifts, or major pricing changes. A model trained on observed sales may forecast sales constrained by inventory availability rather than true customer demand.

Forecast governance should define data ownership, version control, model approval, override authority, monitoring thresholds, and retirement criteria. Fairness concerns may arise when forecasts allocate scarce products across regions or customer groups. A biased historical allocation can become a self-reinforcing forecast if future demand is inferred only from fulfilled orders. Sensitive customer or location data should be minimized and protected.

Modern forecasting platforms may automate parameter estimation, anomaly detection, model selection, and forecast reconciliation across product hierarchies. Machine-learning systems can add value when many explanatory variables and related series are available, but they do not remove the need for leakage controls, temporal validation, interpretability, and human accountability. AI-generated forecasts should expose data provenance, prediction horizon, uncertainty, and override history. Complexity is justified only when it produces measurable decision value.

Best practices include:

• benchmark every model against naïve and seasonal-naïve alternatives;
• validate using future-like splits and multiple origins;
• inspect residual bias and autocorrelation;
• model promotions, stockouts, and calendar effects explicitly;
• use damped trends when indefinite growth is implausible;
• select error measures that reflect operational cost;
• reconcile forecasts across product, location, and aggregate levels when needed;
• communicate intervals and scenarios rather than a falsely precise point estimate;
• monitor performance after implementation and retrain when the process changes.

Chapter Summary

Time-series forecasting uses ordered historical observations to estimate future demand. Naïve methods provide transparent benchmarks; moving averages smooth noise but can lag; exponential smoothing updates a latent level; Holt’s method adds trend; and Holt–Winters adds additive or multiplicative seasonality. Decomposition helps diagnose level, trend, seasonal, and irregular components. ARIMA models represent dependence in stationary or differenced data and require attention to orders, residuals, and stationarity. Models must be compared through chronological validation using appropriate error measures, bias checks, and operational consequences. In supply chains, data quality, stockout treatment, promotions, uncertainty, governance, and structural change are as important as mathematical sophistication.

Key Terms

Autocorrelation: Dependence between observations or errors at different time lags.

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

Bias: Systematic tendency to overforecast or underforecast.

Damped trend: A trend forecast whose future trend contribution gradually weakens.

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

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

Forecast horizon: Number of future periods being forecast.

Holt’s method: Exponential smoothing model with level and trend states.

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

Level: Estimated current baseline of a time series.

MAE: Mean absolute error, expressed in the original demand units.

MAPE: Mean absolute percentage error, unsuitable when actual values are zero or near zero.

Moving average: Mean of a specified number of recent observations.

Naïve forecast: Forecast based on the latest observation or corresponding seasonal observation.

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

Seasonality: Repeating variation associated with a calendar or operational cycle.

Stationarity: Stability of relevant statistical properties over time.

Trend: Persistent directional movement in the series.

Validation: Evaluation of a model on observations not used for fitting.

Review Questions

Short Answer Questions

1. What is the difference between a forecast and an operational plan?
2. When is a seasonal naïve method preferable to a basic naïve method?
3. How does increasing [[M:\alpha]] affect simple exponential smoothing?
4. Distinguish additive from multiplicative seasonality.
5. Why can MAPE be problematic for intermittent demand?

Long Answer Questions

1. Explain the assumptions, strengths, and limitations of moving averages, SES, and Holt’s method.
2. Describe how additive and multiplicative Holt–Winters models update level, trend, and seasonality.
3. Explain a chronological model-validation process for a supply chain forecast.
4. Discuss the meaning of ARIMA([[M:p,d,q]]) and the role of differencing.
5. Evaluate the managerial and governance issues involved in deploying an automated forecasting system.

Analytical / Application Questions

1. A distributor has monthly demand of 210, 230, 250, and 270 units in its latest four months. Calculate the four-month moving-average forecast and explain one limitation of using it for this series.
2. A stable product has seasonal demand that is 40% above baseline every December, while the baseline is increasing. Which forecasting model would you initially test, and why?
3. A model has lower RMSE but substantially greater underforecast bias than a competitor. How should a supply chain manager decide between them when stockout costs are high?

Multiple Choice Questions with Answers

Question 1

Which forecast is produced by the basic naïve method?

A. The mean of all historical observations

B. The most recent observed value

C. The fitted linear trend value

D. The average of the same calendar month across all years

Correct Answer: B. The most recent observed value

Question 2

A four-period moving average uses demands of 80, 100, 120, and 140 units. What is the next forecast?

A. 105 units

B. 110 units

C. 115 units

D. 120 units

Correct Answer: B. 110 units

Question 3

In simple exponential smoothing, a high value of α generally makes the forecast:

A. More responsive to recent observations

B. Less responsive to recent observations

C. Automatically seasonal

D. Independent of the latest observation

Correct Answer: A. More responsive to recent observations

Question 4

Which method explicitly estimates both level and trend but not seasonality?

A. Simple exponential smoothing

B. Holt’s linear trend method

C. Seasonal naïve forecasting

D. Moving-average decomposition

Correct Answer: B. Holt’s linear trend method

Question 5

Multiplicative seasonality is most appropriate when seasonal fluctuations:

A. Have constant absolute size

B. Are unrelated to the demand level

C. Grow or shrink in proportion to the demand level

D. Occur only once

Correct Answer: C. Grow or shrink in proportion to the demand level

Question 6

What is a key limitation of MAPE?

A. It is expressed in demand units

B. It cannot compare forecasts at different scales

C. It is undefined when actual demand is zero

D. It always penalizes large errors more than RMSE

Correct Answer: C. It is undefined when actual demand is zero

Question 7

In ARIMA(p,d,q), the parameter d represents:

A. The seasonal period

B. The number of differencing operations

C. The number of autoregressive terms only

D. The forecast horizon

Correct Answer: B. The number of differencing operations

Question 8

Why should time-series validation generally preserve chronological order?

A. To ensure future information is not used to predict the past

B. To maximize the training sample randomly

C. To eliminate all forecast error

D. To guarantee normally distributed demand

Correct Answer: A. To ensure future information is not used to predict the past

Question 9

A damped trend is useful primarily because it:

A. Removes all seasonal effects

B. Prevents any response to new observations

C. Allows trend influence to weaken over longer horizons

D. Requires zero-valued observations

Correct Answer: C. Allows trend influence to weaken over longer horizons

Question 10

If actual demand is 150 and forecast demand is 120, the forecast error defined as actual minus forecast is:

A. −30 units

B. 0.80 units

C. 30 units

D. 270 units

Correct Answer: C. 30 units