Time-Series Foundations and Benchmark Models

Introduction

Operations managers make decisions in time: how much inventory to hold, how many employees to schedule, how much capacity to reserve, and when to replenish materials. These decisions require forecasts of future demand, lead times, prices, failures, or workloads. A time series is an ordered sequence of observations recorded at successive points in time. Unlike ordinary cross-sectional data, time-series observations are often dependent: sales this week may resemble sales last week, and demand in December may systematically differ from demand in July.

This chapter develops the foundations needed to analyse such dependence and construct defensible benchmark forecasts. The emphasis is not on finding the most sophisticated model immediately. It is on understanding the data-generating structure, identifying level, trend, seasonality and cycles, diagnosing autocorrelation and stationarity, recognizing structural breaks, and establishing simple benchmarks against which advanced models must be judged.

A benchmark is not a naive substitute for analysis. It is a reference standard. If a complex forecasting system cannot outperform a seasonal-naïve forecast after accounting for model complexity and operational cost, its adoption is difficult to justify.

Learning objectives

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

1. Distinguish level, trend, seasonality, cycles, irregular variation and structural breaks.
2. Explain autocorrelation and stationarity and assess their implications for forecasting.
3. Construct naïve, seasonal-naïve, moving-average and decomposition forecasts.
4. Select an appropriate benchmark for a particular operational setting.
5. Evaluate benchmarks using time-ordered validation and interpretable error measures.
6. Diagnose when a benchmark is unreliable because of changing regimes, data problems or inappropriate seasonal assumptions.
7. Communicate forecast assumptions and limitations responsibly.

1. The anatomy of a time series

Let y_t denote the observed value of a variable at time t, where the time index may represent an hour, day, week, month or quarter. A useful conceptual representation is:

Here L_t represents the underlying level, T_t a trend component, S_t a seasonal component, C_t a cyclical component and I_t irregular or unstructured variation. This additive representation is a conceptual decomposition rather than a universal law. It is appropriate when component effects are measured in the same units and seasonal variation is approximately constant in magnitude.

A multiplicative representation is often more suitable when seasonal fluctuations grow with the level:

The multiplicative form requires non-zero, generally positive observations. Taking logarithms converts it into an additive form when all values are positive:

The distinction is operationally important. If a retailer sells approximately 1,000 units in the off-season and 1,500 in the peak season in one year, but 2,000 and 3,000 respectively several years later, the seasonal amplitude has increased with the level. A multiplicative or log-scale treatment may be more realistic than an additive one.

Level

The level is the central or typical magnitude around which observations fluctuate at a particular time. It may change gradually or abruptly. A level is not necessarily the arithmetic mean of the entire historical series, especially when trend or breaks are present.

Trend

Trend is a persistent long-term direction, such as increasing online orders, declining newspaper circulation or gradual growth in installed equipment. Trend need not be linear. A linear trend model can be written as:

where a is the intercept, b is the change per time period, t is the time index and ε_t is the unexplained error. The units of b are units of the variable per time period. A fitted trend should not be extrapolated indefinitely without considering saturation, competition, capacity constraints or changes in the business environment.

Seasonality

Seasonality is a systematic and calendar-related pattern that repeats with a known frequency. Examples include higher electricity demand on weekdays, holiday retail peaks, monthly insurance claims and weekly restaurant demand. If the seasonal period is m, then observations separated by m periods are potential seasonal counterparts. A seasonal pattern may be additive, with a fixed unit effect, or multiplicative, with a proportionate effect.

Seasonality must be distinguished from calendar effects. A month can contain different numbers of working days; Easter can move between calendar months; and public holidays may not occur at fixed positions. Such effects may require calendar regressors rather than a simple fixed seasonal index.

Cycles

Cycles are longer-term rises and falls without a fixed, known period. Business cycles, commodity cycles and equipment replacement cycles are examples. A cycle differs from seasonality because its duration and phase are not reliably fixed. With limited historical observations, cycles are difficult to estimate and forecast. A model that treats a cycle as seasonality can produce false precision.

Irregular variation and outliers

Irregular variation includes random shocks, measurement error and effects not explained by the systematic components. An outlier is an unusually large or small observation. It may represent a data error, a one-time event, or a genuine operational shock. Replacing an outlier automatically is dangerous: a major promotion, strike or weather event may contain information relevant to future planning.

2. Dependence, autocorrelation and stationarity

Autocorrelation

Autocorrelation measures association between observations and their lagged values. The lag-k autocorrelation is commonly defined as:

where Cov is covariance and Var is variance. This ratio is dimensionless and, under standard conditions, lies between −1 and 1. The definition assumes a covariance structure that is stable over the time interval being analysed. In practice, autocorrelation is estimated from a finite sample and is subject to sampling uncertainty.

Positive autocorrelation means that high observations tend to follow high observations and low observations tend to follow low observations. Negative autocorrelation indicates alternation. Strong autocorrelation can make ordinary random-sampling assumptions inappropriate because observations are not independent.

The autocorrelation function, or ACF, displays autocorrelation for several lags. A slowly declining ACF may indicate persistence, trend or non-stationarity. Peaks at seasonal lags may indicate seasonality. The partial autocorrelation function can help distinguish direct dependence at a lag from dependence transmitted through shorter lags, although interpretation should be combined with subject-matter knowledge rather than used mechanically.

Stationarity

A weakly stationary series has a constant mean, constant finite variance and autocovariance that depends on the lag rather than on calendar time. In symbolic form, for all relevant t and k:

Here μ and σ² are constant mean and variance, and γ_k depends only on lag k. Stationarity is a modelling condition, not a claim that every observation is identical. A stationary series may still contain random fluctuations and autocorrelation.

Trend and changing variance often violate stationarity. Differencing can sometimes remove a stochastic trend:

For seasonal differencing with period m:

Differencing should be used because the data and forecasting objective justify it, not simply because a formal test produces a particular result. Over-differencing can create unnecessary noise and induce negative autocorrelation. Unit-root tests and stationarity tests have limited power and should be interpreted alongside plots, domain knowledge and forecast performance.

Structural breaks

A structural break is a persistent change in the data-generating process. Examples include a new competitor, a pricing policy, a pandemic, a regulatory change, a plant relocation or a major technology adoption. A simple mean-shift representation is:

where τ is the break date and μ₁ and μ₂ are different regime means. A break can also affect trend, variance, seasonal amplitude or autocorrelation.

Structural breaks are especially important for benchmarks based on long historical windows. A long moving average may be stable but slow to adapt after a break. A short moving average adapts more quickly but is more variable. The correct choice depends on the cost of forecast errors and the likelihood of regime change.

3. Exploratory analysis before forecasting

A defensible forecasting workflow begins with data understanding rather than model fitting. The analyst should:

1. Confirm the time stamp, frequency, units and aggregation rule.
2. Check missing observations, duplicate records and changes in measurement definitions.
3. Plot the series over time, marking promotions, outages, policy changes and unusual events.
4. Compare the mean and variance across subperiods.
5. Examine seasonal subseries, lag plots and ACF values.
6. Identify whether the forecast target is demand, sales, shipments or another proxy.
7. Establish the operational forecast horizon and decision lead time.

Aggregation can alter the apparent structure. Hourly data may show intraday seasonality, while daily aggregation can conceal it. Aggregating intermittent demand may reduce zeros but also remove useful information. The selected frequency should correspond to the decision being supported.

A decomposition plot is useful because it separates the observed series into interpretable elements. However, decomposition is not proof that the components are uniquely identified. Trend and cycle can be confounded, and a changing seasonal pattern may not be represented by fixed seasonal indices.

4. Benchmark forecasting methods

4.1 Naïve forecast

The naïve method uses the most recent observation as the forecast for every future period:

Here T is the last observed time, h is the forecast horizon, H is the maximum horizon and ŷ_T+h|T is the forecast made at time T. The method assumes that the latest level is the best simple estimate of the near future. It is often strong for short-horizon series with a changing level and little stable seasonality.

The naïve method is also called a random-walk benchmark in some settings, particularly when the underlying stochastic model is y_t = y_t-1 + ε_t. That interpretation requires additional assumptions about the error process; the forecasting rule itself can be used without claiming that the full random-walk model is true.

4.2 Seasonal-naïve forecast

The seasonal-naïve method repeats the observation from the corresponding position in the previous season:

where m is the seasonal period. For monthly data with annual seasonality, m = 12. For daily data with a weekly pattern, m = 7, subject to the treatment of holidays and calendar changes.

This method is a particularly important benchmark when seasonality is stable. It preserves the observed seasonal shape without estimating many parameters. It can fail when the seasonal pattern changes, when the previous season was affected by an exceptional event, or when the horizon extends beyond available seasonal lags.

4.3 Moving-average forecast

A simple moving average of window length k is:

The corresponding one-step forecast is often written:

All observations in the window receive equal weight. The method smooths random variation but introduces lag when the level is changing. For multi-step forecasts, a common simple rule is to use the latest moving average for each future period, although other recursive procedures are possible and must be stated explicitly.

A weighted moving average gives greater weight to selected observations:

with w_i ≥ 0 and ∑w_i = 1. The weights must sum to one so that the forecast retains the same units as y. Weight selection should be validated rather than chosen solely because it looks intuitive.

Worked example: naïve and moving-average forecasts

Suppose weekly demand for the last five weeks is 120, 135, 128, 142 and 150 units.

The naïve forecast for week 6 is the week 5 observation:

For a three-week moving average:

The naïve forecast reacts fully to the latest increase, whereas the moving average smooths the recent values. Neither is automatically superior. If the increase reflects a permanent level shift, the moving average may lag; if the increase is noise, the moving average may be more stable.

4.4 Drift benchmark

When a series has a persistent direction, a drift forecast extends the average historical change:

This rule requires at least two observations and assumes that the average change over the historical interval is informative for the forecast horizon. It is not appropriate when a structural break makes the full historical average obsolete.

4.5 Decomposition forecasts

A decomposition forecast first estimates components and then recombines them. In an additive model:

where T_t is trend, S_t is seasonal effect and R_t is remainder. A future forecast is:

For an ordinary forecast, the expected remainder is usually set to zero, giving:

Seasonal effects in an additive decomposition are commonly constrained to sum to zero over one complete season. For a multiplicative decomposition:

and the forecast is typically:

with expected remainder equal to one. The units are consistent: additive components have the units of the original series, while multiplicative seasonal indices are dimensionless.

A classical decomposition may estimate a moving-average trend, calculate detrended seasonal ratios or differences, average them by season, and project the trend. It is transparent but has limitations: endpoints are difficult to estimate, seasonal effects may be assumed constant, and changing patterns are handled poorly. Modern decomposition procedures can use robust and locally adaptive smoothing, but the same principle remains: decomposition is a representation that must be evaluated by out-of-sample forecasts.

Worked example: additive seasonal decomposition

Assume a monthly service centre has an estimated trend of 1,000 calls in April. Historical analysis indicates additive seasonal effects of −80 for April, +120 for May and +60 for June. If the trend forecast for May is 1,020 calls, the May forecast is:

If the trend forecast for June is 1,040 calls:

These calculations assume that the seasonal effects remain applicable and that the trend estimate is not distorted by a structural break.

5. Forecast evaluation and benchmark selection

Forecast accuracy must be evaluated using simulated real-time information. Randomly splitting observations into training and test sets can leak future information into the past. Instead, use a time-ordered holdout or rolling-origin evaluation. In rolling-origin evaluation, the model is repeatedly trained using data available at an origin and tested on subsequent observations.

Common error measures include:

Mean absolute error:

Root mean squared error:

Mean absolute percentage error is:

MAE and RMSE are measured in the units of the target. RMSE gives greater influence to large errors because errors are squared. MAPE is undefined when y_t = 0 and unstable when actual values are close to zero. For intermittent demand, scaled or absolute measures are generally safer.

A scaled measure compares forecast error with a naïve benchmark. One common form is:

The denominator is the in-sample one-step naïve MAE, and the numerator is the model’s test-set MAE. A value below 1 indicates improvement over that benchmark under the stated evaluation design. For seasonal data, the denominator can instead use seasonal-naïve differences, such as |y_t − y_t-m|.

Accuracy is not the only criterion. A method may be preferred because it is transparent, fast, robust, easy to maintain, or aligned with inventory costs. Bias also matters. Mean forecast error can be monitored as:

A positive value indicates systematic under-forecasting when actuals exceed forecasts on average.

6. Illustrative case: regional pharmacy replenishment

Background

This illustrative case concerns a regional pharmacy chain forecasting weekly demand for a cold-and-flu product. Each store places replenishment orders every week, and the distribution centre needs a four-week forecast. Demand is strongly seasonal, but a public-health campaign and an unusually mild winter may alter the pattern.

Problem

The supply-planning team has been using a six-week moving average. Planners report stockouts during the winter peak and excess inventory after the peak. Management is considering an advanced machine-learning model but first requires a defensible benchmark comparison.

Analysis

The analyst verifies that the series contains weekly store-level demand rather than shipments, because shipment data were constrained by prior stockouts. A plot shows a repeated winter peak and a recent upward shift following the campaign. The ACF has positive values at short lags and a visible peak around 52 weeks, suggesting annual seasonality in weekly data. However, only three complete years are available, so estimates of annual seasonal effects are uncertain.

The following candidates are established:

The evaluation uses rolling four-week forecasts over the most recent winter and non-winter periods. The seasonal-naïve method performs best during stable weeks but under-forecasts after the campaign. The moving average adapts somewhat more quickly but misses the sharp peak. A recent-level naïve benchmark performs well immediately after the campaign but poorly when seasonality returns.

Approach and solution

The team adopts a segmented benchmark policy. The seasonal-naïve forecast remains the baseline for stable stores with sufficient history. A recent-level benchmark is used as a challenger after a verified demand shift. Forecasts are reconciled with known promotional calendars, and planners review exceptions rather than manually replacing every forecast. The evaluation is repeated monthly, with separate reporting for peak and non-peak periods.

Outcome

The principal improvement comes not from selecting one universally superior model but from identifying when each benchmark is credible. The process reduces unexamined reliance on the six-week average and makes model performance visible by regime and horizon.

Learning points

1. The target must represent unconstrained demand as closely as possible.
2. Seasonal-naïve forecasts are powerful baselines when seasonality is stable.
3. Structural breaks can reverse benchmark rankings.
4. Evaluation should reflect the operational horizon and decision cycle.
5. A transparent benchmark policy can coexist with more advanced models.

7. Managerial, governance and technology implications

Forecasts influence purchasing, staffing and service availability, so errors have financial and human consequences. Managers should document the forecast target, time zone, aggregation, horizon, benchmark, training window and treatment of missing or exceptional observations. Every override should have a reason code where practical; otherwise, the organization cannot learn whether human intervention improved or degraded performance.

Forecast governance should include version control, reproducible data preparation, access controls and monitoring for bias across stores, regions or customer groups. A forecast that systematically underestimates demand in a low-volume region may create poorer service there even if aggregate accuracy appears acceptable.

AI and machine learning can identify nonlinear effects, interactions and changing patterns, but they do not remove the need for benchmarks. Automated systems may amplify data leakage, encode historical stockout bias or react poorly to unprecedented events. A strong implementation compares an advanced model with naïve and seasonal-naïve alternatives, uses time-aware validation, monitors drift and preserves an escalation path for exceptional conditions.

8. Best-practice checklist

Before approving a benchmark forecast, ask:

• Is the time index complete and correctly ordered?
• Are units, aggregation and forecast horizon aligned with the decision?
• Has the series been plotted and checked for level, trend, seasonality, cycles and breaks?
• Is the seasonal period justified by operations or merely assumed?
• Has the benchmark been evaluated using information available at the forecast origin?
• Are zero values, outliers and stockout-censored observations treated appropriately?
• Are accuracy, bias, stability and operational cost considered together?
• Are assumptions, overrides and limitations documented?
• Is performance monitored after implementation?

Chapter Summary

Time-series forecasting begins with understanding how observations evolve through time. Level describes the typical magnitude, trend describes persistent direction, seasonality describes systematic calendar repetition, cycles describe less predictable long-term fluctuations, and irregular variation captures residual shocks. Structural breaks represent changes in the underlying data-generating process and can make historical relationships unreliable.

Autocorrelation reveals dependence across lags. Stationarity requires stable mean, variance and lag-based covariance, although practical forecasting decisions should combine statistical evidence with operational knowledge. Differencing can address some forms of non-stationarity but may be harmful when used mechanically.

Naïve, seasonal-naïve, moving-average, drift and decomposition forecasts provide transparent benchmarks. Naïve forecasts are responsive to the latest level; seasonal-naïve forecasts preserve recurring seasonal positions; moving averages smooth noise at the cost of lag; and decomposition forecasts combine estimated trend and seasonal components. No benchmark is universally best. Selection should use rolling or time-ordered evaluation, appropriate error measures and awareness of breaks, calendar effects and business costs.

A benchmark is defensible when its target, assumptions, data treatment, validation design and limitations are explicit. Advanced analytics should demonstrate incremental value over these simple standards rather than replace them without evidence.

Key Terms

Time series: An ordered sequence of observations indexed by time.

Level: The typical magnitude of a series at a given time.

Trend: A persistent long-term direction in a series.

Seasonality: A systematic pattern repeating at a known calendar frequency.

Cycle: A longer-term fluctuation without a fixed, known period.

Irregular component: Unstructured variation not explained by systematic components.

Autocorrelation: Association between observations and their lagged values.

Lag: The number of time periods separating two observations.

ACF: Autocorrelation function showing dependence at successive lags.

Stationarity: Stability of mean, variance and autocovariance over time.

Differencing: Subtracting a previous observation to reduce non-stationarity.

Structural break: A persistent change in the data-generating process.

Naïve forecast: A forecast equal to the most recent observation.

Seasonal-naïve forecast: A forecast equal to the observation from the corresponding prior season.

Moving average: An average of observations in a fixed recent window.

Decomposition: Representation of a series through components such as trend and seasonality.

MAE: Mean absolute error, measured in target units.

RMSE: Root mean squared error, emphasizing large errors.

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

Rolling-origin evaluation: Repeated time-ordered testing that simulates real forecasting.

Review Questions

Short Answer Questions

1. Distinguish seasonality from a cycle.
2. What does positive autocorrelation imply about a time series?
3. State the main assumption of a naïve forecast.
4. Why can MAPE be problematic for intermittent demand?
5. What is a structural break?

Long Answer Questions

1. Explain the additive and multiplicative decomposition frameworks and identify when each is appropriate.
2. Discuss stationarity, including its practical importance and the role of differencing.
3. Compare naïve, seasonal-naïve and moving-average forecasts in terms of assumptions, strengths and limitations.
4. Explain how rolling-origin evaluation should be designed for a four-period operational forecast.
5. Discuss why an advanced forecasting model should be compared with simple benchmarks before deployment.

Analytical / Application Questions

1. A monthly demand series has a strong annual pattern but experienced a permanent price change three months ago. Which benchmarks would you construct, and how would you evaluate them?
2. Weekly demand for the last three weeks is 80, 95 and 100 units. Calculate the one-week-ahead naïve forecast and the three-week moving-average forecast. Explain why the results differ.
3. A forecast system has low RMSE but persistent under-forecasting during peak periods. What additional diagnostics and managerial actions should be considered?

Multiple Choice Questions with Answers

Question 1

Which feature most clearly distinguishes seasonality from a cycle?

A. Seasonality repeats at a known frequency

B. Seasonality is always caused by random shocks

C. A cycle must have a period of exactly 12 months

D. A cycle cannot affect business demand

Correct Answer: A. Seasonality repeats at a known frequency

Question 2

A naïve one-step forecast for time T+1 is defined as:

A. The average of all historical observations

B. The observation at time T

C. The observation at time T−m

D. The fitted long-run trend value

Correct Answer: B. The observation at time T

Question 3

For monthly data with stable annual seasonality, which seasonal period is generally appropriate?

A. 4

B. 7

C. 12

D. 52

Correct Answer: C. 12

Question 4

What is a principal limitation of a simple moving-average forecast when demand is rising?

A. It always produces negative forecasts

B. It ignores all historical observations

C. It can lag behind the changing level

D. It cannot be calculated from numerical data

Correct Answer: C. It can lag behind the changing level

Question 5

Which condition is part of weak stationarity?

A. The mean changes every period

B. The variance is constant over time

C. All observations are equal

D. There is no autocorrelation at any lag

Correct Answer: B. The variance is constant over time

Question 6

If weekly demand is 128, 142 and 150 units in the latest three weeks, what is the three-week moving-average forecast?

A. 128 units

B. 140 units

C. 142 units

D. 150 units

Correct Answer: B. 140 units

Question 7

Why is MAPE unsuitable when an actual observation equals zero?

A. It cannot be calculated because of division by zero

B. It always becomes exactly 100%

C. It ignores the forecast value

D. It changes the units from percentages to units

Correct Answer: A. It cannot be calculated because of division by zero

Question 8

Which event is most likely to create a structural break?

A. Ordinary random variation around a stable level

B. A permanent change in pricing policy

C. A routine seasonal weekend effect

D. Using a shorter moving-average window

Correct Answer: B. A permanent change in pricing policy

Question 9

What does a positive mean forecast error, defined as actual minus forecast, generally indicate?

A. Systematic over-forecasting

B. Systematic under-forecasting

C. Perfect calibration

D. Absence of seasonality

Correct Answer: B. Systematic under-forecasting

Question 10

Why should forecast evaluation generally preserve time order?

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

B. To guarantee that RMSE is zero

C. To eliminate all autocorrelation

D. To make every model seasonal

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