Forecasting and Time-Series Analysis

Introduction

Forecasting is the systematic estimation of future values using historical observations, explanatory information, expert judgement, or a combination of these sources. In business administration, forecasts support capacity planning, budgeting, inventory control, staffing, marketing, financial management, and strategic decision-making. A forecast is not a prediction of certainty; it is a quantified statement about plausible future conditions under specified assumptions.

Time-series analysis focuses on observations recorded sequentially over time. Examples include monthly sales, daily website traffic, quarterly GDP, weekly demand, and hourly electricity consumption. The distinctive feature of time-series data is that observations may be dependent: current demand may reflect previous demand, seasonal routines, promotions, weather, or macroeconomic conditions.

A useful forecasting process begins with a decision, not with a method. The analyst should identify the forecast horizon, frequency, decision consequences, available information, cost of errors, and stability of the underlying environment. A sophisticated model can perform poorly when the data are poorly defined, the horizon is inappropriate, or structural change is ignored.

Learning Objectives

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

1. Explain the components and characteristics of time-series data.
2. Apply moving averages and exponential-smoothing methods.
3. Model trend and seasonality using decomposition and regression.
4. Use regression forecasting while distinguishing association from causation.
5. Explain the logic of introductory ARIMA models.
6. Compare forecasts using appropriate error measures and validation procedures.
7. Evaluate forecasting scenarios, risks, ethics, and managerial implications.

1. The Structure of Time-Series Data

Let y_t denote the observed value of a variable at time t, where t = 1, 2, ..., T. A time series may contain several elements:

• Level: the typical magnitude of the series.
• Trend: persistent long-term increase or decrease.
• Seasonality: a repeating pattern linked to a calendar or operating cycle, such as month, weekday, or quarter.
• Cycle: longer, often less regular fluctuations associated with economic or business conditions.
• Irregular component: random variation, unusual events, measurement error, or shocks.

A classical additive decomposition is:

where T_t is trend, S_t is seasonal variation, and I_t is the irregular component. This form is appropriate when seasonal fluctuations have approximately constant absolute size. A multiplicative decomposition is:

This is more appropriate when seasonal variation grows proportionally with the level. Since multiplication requires positive or appropriately transformed values, additive decomposition is often preferable for series containing zero or negative observations.

Forecasting methods generally assume that the data-generating process will remain sufficiently similar in the forecast period. This assumption is called stability or stationarity, depending on the model. It is never guaranteed. A pandemic, regulatory change, competitor entry, product launch, or change in measurement can make historical relationships unreliable.

Time-Series Plotting and Data Preparation

Before fitting a model, the analyst should plot the series against time and inspect:

• missing observations and irregular time intervals;
• outliers and possible data-entry errors;
• changes in definitions or measurement systems;
• trend, seasonality, and changing variance;
• interventions such as promotions or strikes;
• whether the forecast target is measured in units, currency, rates, or percentages.

Time must be ordered correctly. Randomly shuffling observations before splitting training and test sets causes information from the future to influence model selection and produces optimistic performance estimates.

2. Naïve Forecasts and Benchmarking

A benchmark provides a reference against which more complex models should be judged. The simplest naïve forecast uses the most recent observation:

where F_t+1 is the forecast for the next period. This method is often surprisingly strong for short horizons and highly persistent series.

A seasonal naïve forecast uses the observation from the corresponding previous season:

where m is the seasonal period and h is the forecast horizon. For monthly data with annual seasonality, m = 12. For daily data with a weekly cycle, m = 7.

A forecast should not be called useful merely because it is statistically sophisticated. It should outperform a relevant benchmark on data not used to estimate the model.

3. Moving-Average Forecasting

A simple moving average forecasts the next period by averaging the most recent k observations:

Here, k is the window length. The observations and forecast must use the same units as y_t. A larger window smooths more noise but responds more slowly to changes. A smaller window reacts more quickly but may produce a less stable forecast.

Worked Example: Three-Period Moving Average

Suppose weekly demand for the last three weeks was 120, 135, and 150 units. The forecast for week 4 is:

If actual demand in week 4 is 144 units, a rolling forecast for week 5 using weeks 2–4 is:

The moving average treats all included observations equally. It is therefore less responsive than methods that assign greater weight to recent observations. It also does not naturally model trend or seasonality unless modified or combined with other procedures.

A centered moving average, often used for descriptive decomposition, averages observations before and after a time point. It is useful for estimating the underlying level, but it cannot be used directly for real-time forecasting when future observations are unavailable.

4. Exponential Smoothing

Exponential smoothing assigns the greatest weight to the most recent observation while retaining information from earlier observations through the previous forecast. Simple exponential smoothing is:

where 0 ≤ α ≤ 1 is the smoothing parameter, y_t is the current observation, and F_t is the forecast for period t. A high α reacts rapidly to recent changes; a low α produces greater smoothing. An initial forecast must be specified, commonly as the first observation or an initial average.

Worked Example

Let α = 0.30, current demand y_t = 150, and current forecast F_t = 140. Then:

Simple exponential smoothing is suitable when the series has no systematic trend or seasonality. It is not designed to extrapolate a changing trend effectively.

Holt’s Trend Method

Holt’s method extends exponential smoothing by estimating both level and trend:

Here, ℓ_t is the estimated level, b_t is the estimated trend, β is the trend-smoothing parameter with 0 ≤ β ≤ 1, and h is the number of periods ahead. Holt’s method may be specified with additive or damped trend. A damped trend gradually reduces the extrapolated trend and is often safer for longer horizons because indefinite linear growth is rarely realistic.

Holt–Winters Seasonal Method

When both trend and seasonality are present, Holt–Winters methods estimate level, trend, and seasonal indices. For an additive seasonal model with period m, one common formulation is:

The parameters α, β, and γ lie between 0 and 1. The additive form assumes seasonal effects have roughly constant size. Multiplicative forms use ratios such as y_t ÷ s_t−m and are suitable when seasonal effects scale with the level and observations are positive.

In practice, parameters are often selected by minimizing an error criterion on the training data, but the final choice should be assessed on a validation period. Minimizing in-sample error alone can overfit.

5. Trend and Seasonality Models

Regression with Time Trend

A linear trend model is:

where β₀ is the intercept, β₁ is the average change per time period, and ε_t is the error term. The forecast is the fitted value:

The coefficient b₁ has units of the response per time unit. A linear trend is useful when the relationship is approximately straight over the relevant horizon; it should not be extrapolated mechanically far beyond the observed range.

Seasonal Dummy Variables

For quarterly data, a regression can include three seasonal dummy variables, with one quarter omitted as the reference category:

The omitted quarter is represented by all dummy variables equal to zero. Each δ measures the average difference from the reference quarter, holding trend constant. Including all four quarter dummies along with an intercept creates perfect multicollinearity, known as the dummy-variable trap.

Fourier terms can represent smooth seasonality:

where m is the seasonal period. Additional harmonics can model more complex seasonal shapes, but excessive harmonics may overfit.

Transformations and Multiplicative Effects

If variation increases with the level, a logarithmic transformation may stabilize variance:

A regression fitted to z_t forecasts the logarithm of the outcome. Exponentiating gives a forecast on the original scale, but because E(exp(z)) ≠ exp(E(z)) in general, a bias adjustment may be needed when estimating the conditional mean. Analysts should state whether the objective is a median forecast or a mean forecast on the original scale.

6. Regression Forecasting and Causal Information

Regression forecasting uses explanatory variables such as price, advertising expenditure, income, weather, or interest rates:

The coefficients describe conditional associations under the model. They do not automatically establish causality. Causal interpretation requires a defensible research design and assumptions about confounding, simultaneity, measurement, and model specification.

Regression forecasting is useful when future values of the predictors are known or can themselves be forecast. If future advertising expenditure is unknown, a sales forecast based on advertising requires either a planned advertising scenario or a second forecast for advertising. This creates forecast uncertainty from predictors in addition to uncertainty in the regression relationship.

Time-series regression errors may be autocorrelated. Ordinary least squares coefficient estimates can remain unbiased under some conditions, but standard errors and prediction intervals may be misleading if serial dependence is ignored. Residual plots, autocorrelation diagnostics, robust methods, or dynamic regression models should therefore be considered.

7. Introductory ARIMA Methods

ARIMA models describe dependence within a series after differencing has removed nonstationary trend. The notation ARIMA(p,d,q) contains:

• p: order of the autoregressive component;
• d: number of ordinary differences;
• q: order of the moving-average error component.

A first difference is:

An AR(1) model is:

where φ₁ measures dependence on the previous value. An MA(1) model is:

where the current value depends on the current and previous random shocks. A general ARIMA model applies autoregressive and moving-average terms to the differenced series. Seasonal ARIMA extends this logic with seasonal autoregressive, differencing, and moving-average terms.

A stationary series has a stable probabilistic structure over time, commonly including constant mean and variance and dependence that declines sufficiently with lag. The practitioner may inspect plots, autocorrelation functions (ACF), partial autocorrelation functions (PACF), and unit-root tests. These tools are informative but not infallible; business context and forecast performance remain important.

The Box–Jenkins process is iterative:

1. Identify plausible differencing and model orders.
2. Estimate candidate models.
3. Diagnose residuals.
4. Compare out-of-sample forecasts.
5. Re-estimate and monitor as new data arrive.

A well-specified model should leave residuals that resemble uncorrelated noise. A residual pattern indicates information left unused. However, a statistically significant pattern may have little practical importance, and a visually attractive model may still fail after structural change.

8. Forecast-Error Measures and Model Comparison

Let e_t = y_t − F_t be the forecast error, with actual value y_t and forecast F_t. A positive error means the forecast underestimates the actual value.

Mean Absolute Error

MAE is expressed in the original units and gives equal linear weight to errors of different signs.

Mean Squared Error and Root Mean Squared Error

RMSE is in the original units but penalizes large errors more strongly than MAE.

Mean Absolute Percentage Error

MAPE is interpretable as an average percentage error but is undefined when y_t = 0 and unstable when actual values are close to zero. It can also be asymmetric in managerial interpretation.

Weighted and Scaled Measures

Weighted MAE or weighted RMSE may reflect that errors in high-volume or high-value periods matter more. A scaled measure such as the mean absolute scaled error compares model errors with errors from a naïve benchmark, improving comparability across series. The chosen metric should reflect the decision cost. A retailer facing stockouts may penalize under-forecasting more than over-forecasting; a staffing system may have different costs for each error direction.

Worked Comparison

Suppose actual demand is 100, 120, and 80 units. Model A forecasts 90, 130, and 75; Model B forecasts 105, 110, and 90.

Errors for A are 10, −10, and 5. Thus:

Errors for B are −5, 10, and −10. Thus:

The models tie on MAE. Their RMSE values differ: Model A has √((100+100+25)÷3) ≈ 9.57, while Model B has √((25+100+100)÷3) ≈ 9.57. In this particular example they also tie on RMSE. The result illustrates why model selection should use sufficient test observations and, where relevant, multiple measures.

Forecast evaluation should use rolling-origin or time-series cross-validation. The model is trained on an initial period, forecasts a later period, then the training window expands or rolls forward. This respects temporal order and evaluates the actual operating task.

Prediction intervals communicate uncertainty. A point forecast alone can encourage false precision. Intervals widen with horizon in many models, although the exact form depends on model assumptions and parameter uncertainty.

9. Scenario-Based Forecasting

A scenario is a coherent set of assumptions about future conditions. For example:

• Base case: expected price, demand, and economic conditions.
• Upside case: stronger market growth and successful promotion.
• Downside case: competitor entry, supply disruption, or recession.

Scenario analysis is not the same as assigning arbitrary high and low numbers. Each scenario should identify assumptions, dependencies, probabilities if appropriate, and management actions. Regression models can generate conditional forecasts under alternative predictor values; time-series models can provide a baseline around which scenario adjustments are made.

Decision-makers should distinguish:

• model uncertainty: uncertainty about the correct model;
• parameter uncertainty: uncertainty about estimated coefficients;
• process uncertainty: irreducible variation in future outcomes;
• scenario uncertainty: uncertainty about external conditions.

A forecast dashboard should display the point forecast, interval, key assumptions, recent errors, and triggers for model review.

10. Illustrative Case: Forecasting Demand for a University Café

Background

The fictional Riverside University Café sells prepared meals and beverages. Management must schedule staff and place food orders weekly. Demand has increased over two years, with predictable lower demand during examination breaks and summer closure. Promotions have also produced temporary spikes.

Problem

The manager currently uses the previous week’s sales as the next-week forecast. This creates over-ordering before breaks and stockouts after successful promotions. The manager wants a transparent method that can be updated weekly and compared with alternatives.

Analysis

The analyst first plots 104 weeks of sales and identifies a trend, a strong annual academic-cycle pattern, and several promotion weeks. The data are divided chronologically: the first 80 weeks form the training period, and the final 24 weeks form the test period. A seasonal naïve benchmark uses sales from the corresponding week in the previous academic cycle. Candidate models are:

1. seasonal naïve;
2. four-week moving average;
3. Holt–Winters additive seasonality;
4. regression with time trend, academic-period indicators, and promotion status.

Promotion status is known in advance for planned campaigns. If it is not known, the manager must forecast alternative promotion plans rather than treat the variable as known.

Approach and Solution

The analyst evaluates all models using rolling-origin forecasts and MAE, while separately reporting RMSE because large stockout-producing errors are costly. Residual plots are inspected for remaining seasonality and systematic bias. The regression model performs best during promotion weeks but requires reliable promotion information. Holt–Winters performs more consistently when promotions are unplanned. The seasonal naïve benchmark is competitive during ordinary weeks.

Management therefore adopts a hybrid policy: use the Holt–Winters forecast as the baseline, add a documented promotion adjustment when a campaign is approved, and compare weekly performance against the seasonal naïve benchmark. Forecasts are accompanied by an 80% prediction interval and an escalation rule when actual demand falls outside the interval twice in four weeks.

Outcome

After one academic term, MAE falls relative to the previous-week method, and food waste declines because forecasts are less inflated before breaks. Stockouts remain during unexpected events, but the monitoring rule prompts faster review. The manager does not treat the model as an automatic ordering authority; operational judgement remains necessary for perishables and supplier constraints.

Learning Points

1. A benchmark is essential for demonstrating value.
2. Calendar knowledge can be as important as mathematical complexity.
3. Explanatory variables are useful only when their future values are available or scenario-defined.
4. Error costs should influence evaluation metrics.
5. Monitoring and governance are part of forecasting, not postscript activities.

11. Implementation, Governance, and Professional Practice

A defensible forecasting system should maintain a data dictionary, version-controlled code, documented assumptions, and an audit trail of forecast revisions. Data access should follow authorization and privacy principles. Although sales forecasting usually uses aggregated data, customer-level data may reveal personal or commercially sensitive information.

Automated and AI-assisted forecasting tools can rapidly compare models, detect anomalies, and generate explanations. They do not eliminate the need for human oversight. Risks include leakage of future information, biased training data, opaque transformations, inappropriate extrapolation, and automation bias. A model should be challenged by domain experts, tested on temporally appropriate data, and monitored for drift.

Ethically responsible forecasting requires transparency about uncertainty. Presenting a single precise number when the interval is wide can mislead investors, employees, suppliers, or public decision-makers. Forecasts that influence staffing or credit decisions should be checked for disparate impacts and should permit review of material errors.

Best practice includes:

• define the decision and horizon before choosing a model;
• establish naïve benchmarks;
• preserve temporal ordering in validation;
• compare models using decision-relevant metrics;
• inspect residuals and forecast bias;
• communicate intervals and assumptions;
• monitor performance after deployment;
• review the model after structural breaks;
• retain human accountability for consequential decisions.

Chapter Summary

Forecasting converts historical and contextual information into structured estimates of future conditions. Time-series observations may contain level, trend, seasonality, cycles, and irregular variation. Moving averages provide transparent smoothing; exponential smoothing adapts weights toward recent observations; Holt and Holt–Winters methods extend smoothing to trend and seasonality. Regression models incorporate explanatory variables but require future predictor values and should not be interpreted causally without appropriate assumptions. ARIMA models represent dependence in differenced series and are developed through identification, estimation, diagnosis, and validation.

Forecast quality must be assessed out of sample using measures such as MAE, RMSE, MSE, and carefully interpreted percentage measures. Rolling-origin validation respects temporal order. Scenario analysis makes external assumptions explicit, while prediction intervals communicate uncertainty. Effective forecasting is therefore not simply a contest for the lowest error: it is a governed managerial process linking data, models, decisions, costs, and changing conditions.

Key Terms

Time series: Observations recorded sequentially over time.

Forecast: An estimate of a future observation or condition.

Level: The underlying typical magnitude of a series.

Trend: Persistent long-term movement in a series.

Seasonality: A repeating pattern linked to a fixed calendar or operating period.

Cycle: Longer-term fluctuation that is not necessarily fixed or regular.

Irregular component: Unsystematic variation or shock.

Naïve forecast: A forecast based on a recent or corresponding historical observation.

Moving average: An average of a fixed number of recent observations.

Exponential smoothing: A method that weights recent observations more heavily through recursive updating.

Smoothing parameter: A value controlling responsiveness to new observations.

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

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

Autocorrelation: Dependence between observations separated by a specified lag.

Stationarity: A stable statistical structure over time.

ARIMA: Autoregressive integrated moving-average model.

MAE: Mean absolute error, measured in original units.

RMSE: Root mean squared error, which penalizes large errors more heavily.

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

Scenario analysis: Evaluation of forecasts under coherent alternative assumptions.

Review Questions

5 Short Answer Questions

1. Distinguish trend from seasonality in a time series.
2. What is the main effect of increasing the moving-average window length?
3. State the simple exponential-smoothing equation and define its parameter.
4. Why must one seasonal dummy variable be omitted when an intercept is included?
5. Why is a naïve forecast an important benchmark?

5 Long Answer Questions

1. Explain the classical additive and multiplicative decomposition models and identify when each is appropriate.
2. Compare simple exponential smoothing, Holt’s method, and Holt–Winters methods in terms of data patterns and assumptions.
3. Explain how regression forecasting differs from causal analysis.
4. Describe the Box–Jenkins process for developing an introductory ARIMA model.
5. Discuss how forecast-error measures can influence managerial model selection.

3 Analytical / Application Questions

1. A retailer has monthly sales with an upward trend and strong annual seasonality. Recommend two candidate forecasting methods, explain your choice, and design a time-ordered validation plan.
2. A firm’s regression forecast uses price and advertising expenditure, but next quarter’s advertising budget has not been approved. Explain how this affects the forecast and propose a scenario framework.
3. Two models have equal MAE, but one has substantially higher RMSE and produces more under-forecasts during stockout-prone periods. Which model should management prefer, and what additional evidence would you request?

Multiple Choice Questions with Answers

Question 1

Which component represents a repeating pattern linked to a calendar period?

A. Trend

B. Seasonality

C. Irregular variation

D. Parameter uncertainty

Correct Answer: B. Seasonality

Question 2

If the last three observations are 120, 135, and 150 units, what is the three-period moving-average forecast?

A. 125 units

B. 130 units

C. 135 units

D. 140 units

Correct Answer: C. 135 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. Independent of the latest observation

D. Automatically seasonal

Correct Answer: A. More responsive to recent observations

Question 4

Which method is specifically designed to estimate level, trend, and seasonality?

A. Simple moving average

B. Simple exponential smoothing

C. Holt–Winters smoothing

D. Naïve forecasting

Correct Answer: C. Holt–Winters smoothing

Question 5

In a regression with an intercept and seasonal dummy variables, one category is normally omitted to avoid:

A. Heteroskedasticity

B. Perfect multicollinearity

C. Underfitting

D. Autocorrelation

Correct Answer: B. Perfect multicollinearity

Question 6

Which measure is expressed in the original units and gives linear weight to absolute errors?

A. MAE

B. MAPE

C. MSE

D. R²

Correct Answer: A. MAE

Question 7

A seasonal naïve forecast for a monthly series with annual seasonality usually uses:

A. The immediately preceding day

B. The average of all historical months

C. The value from the same month in the previous year

D. The fitted regression intercept

Correct Answer: C. The value from the same month in the previous year

Question 8

In ARIMA notation, the letter I refers to:

A. Intervention

B. Integration through differencing

C. Information criteria

D. Independent variables

Correct Answer: B. Integration through differencing

Question 9

Why should time-series validation preserve chronological order?

A. To ensure future information does not enter model training

B. To increase the number of observations

C. To remove all forecast uncertainty

D. To guarantee normally distributed errors

Correct Answer: A. To ensure future information does not enter model training

Question 10

When actual values can be zero, which error measure should not be used without modification because it becomes undefined?

A. MAE

B. RMSE

C. MAPE

D. MSE

Correct Answer: C. MAPE