Time-Series Analysis, Forecasting and Index Numbers

Introduction

Business decisions are often made in time: a retailer plans inventory for next month, a bank estimates future loan demand, and a government monitors inflation over several years. A time series is a sequence of observations recorded in chronological order, usually at equal intervals. Time-series analysis seeks to understand the structure of historical observations and use that structure to support forecasts.

Forecasting does not mean predicting the future with certainty. It means producing a systematic estimate, together with an assessment of likely error. Good forecasting combines statistical methods, business knowledge, data-quality controls and continuous monitoring. This chapter also examines index numbers, which summarize relative changes in prices, quantities or business activity across time or locations.

Learning Objectives

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

• distinguish trend, seasonality, cyclical movement, irregular movement and autocorrelation;
• construct and interpret moving-average and exponential-smoothing forecasts;
• use regression for trend and seasonal forecasting;
• evaluate forecasts using appropriate accuracy measures;
• construct and interpret price, quantity and business index numbers;
• identify assumptions, limitations, ethical issues and practical risks in forecasting.

1. The Structure of a Time Series

Let y_t denote the observed value at time t, where time may represent a day, month, quarter or year. A time series should retain its order because rearranging the observations destroys information about temporal dependence.

A useful conceptual decomposition is:

where T_t is trend, S_t is seasonal movement, C_t is cyclical movement, and I_t is irregular movement. This is an additive decomposition, appropriate when seasonal fluctuations have approximately constant absolute size. A multiplicative decomposition is:

The multiplicative form is appropriate when seasonal variation grows or shrinks with the level of the series. In practice, logarithms can convert a multiplicative relationship into an additive one, provided all observations are positive.

Trend

Trend is the long-term direction of movement. It may be upward, downward or approximately horizontal. Trend can result from population growth, technological change, income growth or changing consumer preferences. A trend is not necessarily linear.

Seasonality

Seasonality is a regularly repeating pattern associated with the calendar or another known cycle. Examples include higher hotel demand during holidays, increased electricity use in summer and lower retail sales immediately after major festivals. Seasonal effects have a known and stable period, such as 12 months or four quarters.

Cyclical movement

Cycles are medium- or long-term fluctuations around trend, often linked to economic expansions and contractions. Unlike seasonality, cycles do not have a fixed length or stable amplitude. Business-cycle effects are therefore more difficult to forecast.

Irregular movement

Irregular movement consists of unpredictable shocks such as strikes, natural disasters, sudden regulatory changes or extraordinary events. Robust forecasting methods should avoid interpreting a one-time shock as a permanent trend.

Autocorrelation

Autocorrelation measures the relationship between observations separated by a lag. For lag k, the sample autocorrelation is commonly estimated by:

Here, n is the number of observations, ȳ is the sample mean, and k is a positive integer lag. The numerator measures paired deviations at distance k, while the denominator standardizes by total variation. Positive autocorrelation means high values tend to follow high values and low values tend to follow low values. Negative autocorrelation indicates alternation.

Autocorrelation matters because many statistical procedures assume independent errors. If a regression model leaves autocorrelated residuals, standard errors and significance tests may be misleading, and forecasts may be inefficient.

2. Exploring and Preparing Time-Series Data

Before selecting a forecasting method, an analyst should:

1. define the forecasting objective and horizon;
2. identify the time interval and ensure observations are correctly ordered;
3. inspect missing, duplicated and anomalous observations;
4. plot the series and, where useful, seasonal subseries plots;
5. examine changes in variance and possible structural breaks;
6. separate information available at the forecast date from information that became available later.

A time plot often reveals more than a table. A rapidly increasing variance may suggest a logarithmic or other variance-stabilizing transformation. Missing observations should not be silently replaced with zeros. Imputation must use a documented method and preserve the distinction between a true zero and an unobserved value.

A forecast evaluation should use a time-ordered split. The training period precedes the validation period, and the test period follows both. Randomly splitting observations can allow information from the future to enter model development.

3. Moving-Average Forecasting

A simple moving average smooths short-term fluctuations. For an m-period moving average, the smoothed value at time t is:

The formula requires at least m consecutive observations. If the moving average is used to forecast the next period, the forecast made at time t is:

All observations receive equal weight. A larger m produces greater smoothing but responds more slowly to recent changes. A smaller m responds more quickly but retains more random variation.

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 165 units, the next forecast is based on 135, 150 and 165:

This method is simple and transparent, but it is unsuitable without adjustment when strong trend or seasonality is present. It also cannot produce a forecast for the first m−1 periods without an initialization rule.

A weighted moving average assigns weights w_i to recent observations:

The usual conditions are w_i ≥ 0 and ∑w_i = 1. These conditions preserve the units and keep the forecast within the weighted range of the observations.

4. Exponential Smoothing

Simple exponential smoothing updates the forecast using the latest forecast error:

where 0 < α < 1 is the smoothing constant, y_t is the latest actual value, and F_t is the forecast for time t. Equivalently:

The term y_t − F_t is the forecast error. A high α gives greater weight to recent data and reacts quickly; a low α produces greater smoothing. Simple exponential smoothing assumes no systematic trend or seasonality. The starting forecast F₁ must be specified, often as the first observation or an initial average.

Worked Example

Let F₁ = 100, α = 0.30, and the actual value at time 1 be 110. Then:

If y₂ = 106, then:

For data with trend, Holt’s method adds a level and trend component. One common form is:

Here, L_t is the estimated level, B_t is the estimated trend, 0 < β < 1 is the trend-smoothing constant, and h is the number of periods ahead. Seasonal data may require Holt–Winters methods, which add a seasonal component. The chosen additive or multiplicative version must match the observed behavior of seasonal variation.

5. Regression Forecasting

Regression expresses a response variable as a function of explanatory variables. A linear time-trend model is:

where β₀ is the intercept, β₁ is the average change per time unit, t is a time index, and ε_t is the error. The fitted forecast is:

The slope estimate in ordinary least squares is:

and the intercept is b₀ = ȳ − b₁t̄. The slope has units of the dependent variable per time unit. A linear trend should not be extrapolated indefinitely without considering market limits, saturation or structural change.

Seasonality can be represented with dummy variables. For quarterly data, one quarter is omitted as the reference category:

Here, Q_1t, Q_2t and Q_3t equal 1 when observation t belongs to the relevant quarter and 0 otherwise. The omitted quarter is represented by the intercept. Each γ measures the seasonal difference from the reference quarter, holding trend constant.

Regression forecasting is useful when external predictors are known or can themselves be forecast, such as advertising expenditure, price, income or weather. However, correlation does not establish causation. A predictor that is unavailable at the forecast date cannot be used operationally unless its own forecast is incorporated.

6. Forecast Accuracy and Model Selection

Let the forecast error be:

The mean absolute error is:

The mean squared error is:

The root mean squared error is:

The mean absolute percentage error is:

MAE and RMSE have the same units as the data. RMSE penalizes large errors more heavily because errors are squared. MAPE is easy to communicate as a percentage but is undefined when y_t = 0 and unstable when actual values are close to zero. In such cases, MAE, RMSE or scale-free alternatives are preferable.

A forecast can be accurate on average but biased. Mean error is:

A positive ME indicates that forecasts tend to be too low, while a negative ME indicates overforecasting, assuming error is defined as actual minus forecast. Accuracy measures should be calculated on observations not used to fit the model. Rolling-origin evaluation, in which the model is repeatedly trained on earlier data and tested on the next period, is particularly appropriate for time series.

Model selection should consider accuracy, interpretability, cost, stability and operational consequences. The model with the lowest historical error is not automatically the best model if it is unstable, impossible to explain or dependent on unavailable inputs.

7. Index Numbers

An index number expresses the relative level of a variable compared with a base. If the base-period value is assigned 100, an index of 125 means the current value is 25% above the base, while an index of 90 means it is 10% below the base.

Simple Price and Quantity Indices

For commodity i, the price relative is:

where p_i0 is the base-period price and p_i1 is the current-period price. The quantity relative is:

A simple aggregate price index is:

It is easy to calculate but ignores the relative importance of commodities and can be distorted by measurement units.

Weighted Price Indices

The Laspeyres price index uses base-period quantities as weights:

The Paasche price index uses current-period quantities:

The Fisher ideal price index is the geometric mean of the two:

The same logic applies to quantity indices. A Laspeyres index tends to use an older consumption basket and may not reflect substitution toward relatively cheaper goods. A Paasche index uses the current basket but requires current-period quantity data. Index construction therefore involves both statistical and judgmental decisions.

Worked Index Example

Suppose a business sells two products:

For the Laspeyres price index:

Using the base-period basket, prices increased by 15%. The Paasche index is:

The Fisher index is approximately:

Thus, the Fisher measure indicates an overall price increase of approximately 14.64% relative to the base period.

Business and Financial Indices

A business index may combine sales, orders, employment, production, customer traffic or confidence. Before combining components, the analyst should define the population, base period, weights, units, revision policy and treatment of missing data. A composite index may be a weighted sum of standardized component indices:

Weights may reflect expenditure shares, revenue importance, employment shares or expert judgment. A changing-base index updates the reference period periodically. A chain index links successive period-to-period changes and can better accommodate changing product composition, but it may be less intuitive and can lose some additive properties.

8. Illustrative Case: Forecasting Retail Demand

Background. A regional grocery chain must forecast weekly demand for a fresh product. Overstock creates spoilage, while understock causes lost sales. Historical data show a gradual upward trend and unusually high demand during holiday weeks.

Problem. The operations team initially uses a three-week moving average. Forecast errors are large during holiday periods because the method ignores seasonality. Managers need a method that is accurate, explainable and available within the existing planning system.

Analysis. The analyst plots weekly demand, identifies holiday indicators, checks missing records and separates the final eight weeks as a test set. Residual inspection shows that the moving-average forecast is systematically low during holidays. A trend-plus-seasonal regression is estimated using a time index and holiday dummy. Holt’s trend method is also evaluated. MAE, RMSE and bias are calculated on the test period.

Approach and solution. The regression model produces lower test MAE and substantially reduces holiday bias. However, the model requires a reliable holiday calendar and an estimate of promotional activity. The chain therefore uses regression as the primary forecast and a simpler exponential-smoothing forecast as a fallback when promotion information is missing. Forecasts are reviewed weekly, and large residuals trigger investigation rather than automatic parameter changes.

Outcome. The business reduces average forecast error and spoilage while maintaining a transparent explanation for store managers. The model is not treated as permanent: performance is monitored as product mix, pricing and customer behavior change.

Learning points. A method should reflect the data-generating structure; evaluation must be out-of-sample; operational availability matters; and human review is valuable for unusual events but should be documented to prevent arbitrary overrides.

9. Limitations, Governance and Contemporary Practice

Forecasts inherit the limitations of their data. Changes in definitions, product assortments, prices, distribution channels or customer behavior can create structural breaks. Historical relationships may fail during crises. Analysts should report prediction intervals where possible, rather than presenting a point forecast as certain.

Ethical practice requires transparent documentation of data sources, transformations, exclusions, model versions, assumptions and limitations. Forecasts used for staffing, credit, pricing or access to services can affect people unequally. Analysts should test for systematic errors across relevant groups and avoid using sensitive variables without a legitimate, documented purpose.

Automated and AI-based forecasting systems can process many series and detect nonlinear patterns. Their use does not remove the need for time-ordered validation, leakage controls, monitoring, explainability and human accountability. A complex model should demonstrate material improvement over a strong baseline before it is adopted. Data privacy, access control, retention and audit trails are essential when forecasts use customer or employee data.

Best practice is to maintain benchmark models, compare several horizons, monitor forecast bias, retrain only when justified, record overrides, and establish thresholds for model review. Index numbers should also be revised transparently when improved data or methods become available.

Chapter Summary

Time-series analysis studies chronologically ordered data and separates, conceptually, trend, seasonality, cycles and irregular variation. Autocorrelation reveals dependence across time and warns analysts that independent-error assumptions may be inappropriate. Moving averages provide transparent smoothing, while exponential smoothing updates forecasts using recent errors and can be extended to trend and seasonality. Regression models incorporate trend, seasonal indicators and external predictors, but their inputs must be available at the forecast date and their residuals should be examined.

Forecast accuracy should be assessed on future-like data using measures such as MAE, MSE, RMSE, MAPE and mean error. No single measure is universally best. Index numbers summarize relative changes and require careful choices about base periods, weights, quantities, aggregation and revisions. Laspeyres, Paasche and Fisher indices use different weighting perspectives. In professional practice, statistical accuracy must be combined with interpretability, governance, ethical review and continuous monitoring.

Key Terms

Time series: Chronologically ordered observations of a variable.

Trend: Long-term direction of movement in a series.

Seasonality: A regularly repeating pattern linked to a calendar or known cycle.

Cyclical movement: Medium- or long-term fluctuation without a fixed period.

Irregular movement: Unpredictable or one-time variation.

Autocorrelation: Correlation between observations separated by a specified lag.

Lag: The number of time intervals between two observations.

Moving average: An average of observations in a rolling time window.

Weighted moving average: A moving average that assigns unequal weights to observations.

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

Smoothing constant: A parameter controlling responsiveness to recent data.

Regression forecast: A forecast generated from an estimated relationship between a response and predictors.

Forecast error: The difference between actual and forecast values.

MAE: Mean absolute error, the average absolute forecast error.

RMSE: Root mean squared error, a measure that penalizes large errors.

Index number: A relative measure comparing a value or aggregate with a base.

Laspeyres index: An index using base-period quantities as weights.

Paasche index: An index using current-period quantities as weights.

Fisher index: The geometric mean of Laspeyres and Paasche indices.

Base period: The reference period assigned an index value, commonly 100.

Review Questions

Short Answer Questions

1. What is the difference between trend and seasonality?
2. Why can autocorrelation create problems in regression analysis?
3. State the formula for a three-period moving-average forecast.
4. What does a smoothing constant control in exponential smoothing?
5. Why is MAPE unsuitable when actual values are zero?

Long Answer Questions

1. Explain additive and multiplicative time-series decomposition and identify when each is appropriate.
2. Compare moving averages, exponential smoothing and regression forecasting in terms of assumptions, responsiveness and interpretability.
3. Explain how time-ordered validation differs from random train-test splitting.
4. Compare Laspeyres, Paasche and Fisher price indices, including their weighting implications.
5. Discuss the governance and ethical requirements for deploying automated business forecasts.

Analytical / Application Questions

1. A retailer records demand of 80, 90 and 100 units in the last three periods. Calculate the next three-period moving-average forecast. If actual demand is 110, calculate the following forecast.
2. Using α = 0.25, F_t = 200 and y_t = 220, calculate F_t+1 and interpret the result.
3. Design a forecasting evaluation plan for a business with strong monthly seasonality. Specify the data split, benchmark model, accuracy measures and monitoring actions.

Multiple Choice Questions with Answers

Question 1

Which component of a time series is associated with a regularly repeating calendar pattern?

A. Trend

B. Seasonality

C. Irregular movement

D. Sampling error

Correct Answer: B. Seasonality

Question 2

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

A. 130 units

B. 135 units

C. 140 units

D. 150 units

Correct Answer: B. 135 units

Question 3

In simple exponential smoothing, increasing α generally makes the forecast:

A. Less responsive to recent observations

B. More responsive to recent observations

C. Unaffected by the latest observation

D. Equal to the long-run trend automatically

Correct Answer: B. More responsive to recent observations

Question 4

If actual demand is 220 and the forecast is 200, with error defined as actual minus forecast, the forecast error is:

A. −20

B. 0

C. 20

D. 420

Correct Answer: C. 20

Question 5

Which accuracy measure penalizes large errors relatively more because errors are squared?

A. MAE

B. MSE

C. MAPE

D. Mean error

Correct Answer: B. MSE

Question 6

Why can random train-test splitting be inappropriate for time-series data?

A. It always reduces the sample size to zero

B. It can allow future information to influence model development

C. It prevents calculation of RMSE

D. It removes all seasonal variation

Correct Answer: B. It can allow future information to influence model development

Question 7

A Laspeyres price index uses which quantities as weights?

A. Base-period quantities

B. Current-period quantities

C. Average quantities only

D. Quantities from a randomly selected period

Correct Answer: A. Base-period quantities

Question 8

An index value of 125 with a base value of 100 indicates that the measured quantity is:

A. 25% below the base

B. Unchanged from the base

C. 25% above the base

D. 125% below the base

Correct Answer: C. 25% above the base

Question 9

What is a key limitation of MAPE?

A. It has no percentage interpretation

B. It cannot be calculated when actual values are zero

C. It always penalizes large errors more than RMSE

D. It requires price data

Correct Answer: B. It cannot be calculated when actual values are zero

Question 10

Which statement best describes autocorrelation?

A. The relationship between two unrelated businesses

B. The relationship between observations separated by a time lag

C. The difference between price and quantity indices

D. The average of all forecast errors

Correct Answer: B. The relationship between observations separated by a time lag