Statistical Foundations and Forecast Uncertainty

Introduction

Demand forecasting is not simply the production of a number. It is the disciplined estimation of future demand from incomplete, variable and potentially biased information. Logistics decisions—inventory positioning, transport capacity, labour scheduling, procurement and service commitments—depend not only on an expected demand value but also on the uncertainty surrounding it.

Statistical foundations provide the language and tools for separating systematic signal from random noise. Descriptive statistics summarise observed demand; probability models represent possible future outcomes; correlation helps identify relationships; sampling supports inference from partial data; and confidence and prediction intervals communicate uncertainty. These tools are particularly important in logistics because forecast errors have asymmetric consequences: excess inventory creates holding and obsolescence costs, whereas shortages may cause lost sales, expediting and reputational damage.

A statistically sophisticated forecast is therefore not necessarily the most useful forecast. The analyst must match the method to the decision, state assumptions clearly, quantify uncertainty honestly and communicate results in managerial terms.

Learning objectives

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

1. Distinguish populations, samples, parameters and statistics.
2. Calculate and interpret measures of centre, dispersion and relative variability.
3. Apply probability concepts and common distributions to logistics demand.
4. Interpret covariance, correlation and regression relationships without confusing association with causation.
5. Explain sampling error, confidence intervals and prediction intervals.
6. Distinguish signal from noise and identify conditions that invalidate statistical conclusions.
7. Communicate forecast uncertainty in a way that supports inventory and capacity decisions.

1. Data, populations and statistical reasoning

A population is the complete set of units or events of interest—for example, all daily customer orders during a year. A sample is a subset used to learn about that population. A population quantity is a parameter; a quantity calculated from a sample is a statistic.

In forecasting, the distinction is temporal. Historical observations are available, but the future demand distribution is unknown. A sample mean may estimate the historical average, yet it does not guarantee that the future average will be identical. Changes in price, market structure, service policy or customer behaviour can make historical data non-representative.

A useful statistical workflow is:

1. Define the decision and the forecast horizon.
2. Define the observational unit and demand measure.
3. Inspect data quality, missingness, outliers and aggregation.
4. Describe the data using appropriate statistics and graphics.
5. Specify a probability or forecasting model.
6. Estimate parameters and quantify uncertainty.
7. Validate assumptions using holdout data and residual diagnostics.
8. Translate results into operational decisions.

The workflow is iterative. A mathematically precise estimate based on poorly defined demand—such as shipments rather than customer orders when stockouts exist—can be operationally misleading.

2. Descriptive statistics for logistics data

Measures of centre

For observations x₁, x₂, …, xₙ, the arithmetic mean is:

Here x̄ is the sample mean, xᵢ is observation i, and n is the number of observations. The units of x̄ are the same as the observations, such as units per day. The mean is sensitive to extreme values.

The median is the middle ordered observation, or the average of the two middle observations when n is even. It is more resistant to outliers and is often informative for skewed order sizes. The mode is the most frequent value and can be useful for discrete case quantities, although a dataset may have several modes or none.

Worked example. Daily orders over five days are 8, 10, 10, 12 and 25 cases. The mean is x̄ = (8+10+10+12+25)/5 = 13 cases. The median is 10 cases. The unusually high day raises the mean above the typical central observation. A manager using only the mean may overstate routine replenishment needs if the 25-case observation was a one-off promotion.

Measures of dispersion

The range is maximum minus minimum. It is easy to understand but depends entirely on two observations. The sample variance is:

The sample standard deviation is:

The denominator n − 1, called Bessel’s correction, is used when s² estimates the variance of a population from a sample under standard random-sampling assumptions. If the data constitute the entire finite population of interest, a population variance with denominator n may be appropriate. Variance is expressed in squared units, whereas standard deviation returns to the original units.

For the five-day example, deviations from 13 are −5, −3, −3, −1 and 12. Their squared deviations sum to 188. Thus s² = 188/4 = 47 cases² and s = √47 ≈ 6.86 cases. The large standard deviation signals substantial variability relative to the mean.

The coefficient of variation (CV) compares dispersion across series with different means:

It is dimensionless and is often reported as a percentage. It is meaningful primarily when the mean is positive and substantively away from zero. For the example, CV = 6.86/13 ≈ 0.528, or approximately 52.8%. A high CV may justify safety stock, flexible capacity or segmentation, but it does not by itself identify the cause of variability.

Percentiles and robust summaries

The pth percentile is a value below which approximately p percent of observations fall. The interquartile range is:

where Q₁ and Q₃ are the 25th and 75th percentiles. The IQR is resistant to extreme observations. In logistics, the 90th or 95th percentile of daily demand may be more relevant than the mean when designing capacity, while the median may better describe a normal operating day.

Descriptive statistics should be accompanied by plots. A time plot reveals trend, seasonality, level shifts and changing variance; a histogram reveals skewness and multimodality; a box plot highlights unusual observations. Aggregation can conceal important behaviour: monthly totals may hide weekday peaks and end-of-month surges.

3. Probability and uncertainty

Probability quantifies uncertainty about events. A random variable represents a numerical outcome whose value is uncertain. A discrete random variable has countable outcomes, such as the number of orders; a continuous random variable can take values on an interval, such as delivery time.

For an event A, the complement rule is P(Aᶜ) = 1 − P(A). For mutually exclusive events A and B, P(A ∪ B) = P(A) + P(B). In general:

Conditional probability is:

Independence means that knowledge of B does not change the probability of A, so P(A | B) = P(A) and P(A ∩ B) = P(A)P(B). Demand and supplier lead time, for example, should not automatically be assumed independent; promotions may increase demand while simultaneously creating supplier delays.

The expected value of a discrete random variable X is:

Its variance is:

where μ = E(X). Expected value is a long-run average, not a promise that the outcome will equal that value.

4. Distributions used in demand analysis

Bernoulli and binomial distributions

A Bernoulli trial has two outcomes, such as whether an order arrives in a specified short interval. If X counts successes in n independent trials with constant success probability p, then X has a binomial distribution:

where k = 0, 1, …, n and C(n,k) is the number of combinations. The assumptions—fixed n, binary outcomes, constant p and independence—must be defensible. They often fail when demand is clustered or affected by promotions.

Poisson distribution

The Poisson model is commonly used for counts in a fixed interval when events occur independently at a stable average rate λ:

For a Poisson variable, both the mean and variance equal λ. If observed variance substantially exceeds the mean, overdispersion may indicate demand heterogeneity, dependence or an unsuitable model. A negative binomial or a hierarchical model may then be more appropriate.

Normal distribution

The normal distribution is continuous, symmetric and described by mean μ and standard deviation σ. Standardisation is:

The normal model is useful for aggregated demand and forecast errors when the distribution is approximately symmetric. It can be inappropriate for intermittent, bounded, strongly skewed or highly promotional demand. A normal approximation may also assign a small probability to negative demand, which is impossible for physical units.

Lognormal and empirical distributions

If the logarithm of a positive variable is approximately normal, the variable may be lognormal. This can represent multiplicative variation in order values or demand volumes. However, selecting a distribution should follow data diagnostics, not convenience. Empirical distributions, bootstrap resampling, quantile regression and simulation can avoid imposing an unrealistic parametric shape.

5. Correlation, covariance and regression

Covariance measures joint variation:

The Pearson correlation is:

where σₓ and σᵧ are positive standard deviations. The sample correlation r lies between −1 and +1. A positive value indicates that high values of one variable tend to accompany high values of the other; a negative value indicates the opposite. Correlation measures linear association, not causality. It can be distorted by outliers, common trends, range restriction and aggregation.

A simple linear regression model is:

Here yᵢ is the response, xᵢ is a predictor, β₀ is the intercept, β₁ is the expected change in y for a one-unit increase in x, and εᵢ is an error term. Least squares estimates β₀ and β₁ by minimising the sum of squared residuals. Interpretation requires attention to omitted variables, reverse causality, nonlinearity and extrapolation.

For logistics, temperature, price, advertising, calendar effects and service levels may explain demand. A strong correlation between promotions and demand does not prove that every promotion causes the same increase: promotions may be targeted precisely when demand is expected to be high.

6. Sampling and the central limit theorem

Sampling is unavoidable when data collection is costly or when the population is conceptually future demand. A probability sample gives each population unit a known, non-zero selection probability. Simple random sampling, stratified sampling and systematic sampling have different advantages. Convenience samples are fast but can be severely biased.

The standard error of the sample mean under independent observations with finite variance is:

When σ is unknown, it is estimated by s. Increasing n reduces sampling variability at a rate proportional to 1/√n; quadrupling sample size approximately halves the standard error. This does not remove systematic bias.

The central limit theorem states, under broad conditions, that the sampling distribution of the mean approaches normality as n increases, even when individual observations are not normal. The approximation is less reliable for very small samples, extreme skewness, dependence or infinite-variance processes. Time-series observations often violate independence, so effective sample size may be much smaller than the number of recorded periods.

7. Confidence intervals and prediction intervals

A confidence interval estimates an unknown population parameter. For a mean with known σ, a 100(1 − α)% interval is:

where z₁₋α/₂ is the relevant standard-normal quantile. When σ is unknown and the sample is reasonably random, independent and approximately normal, use the t distribution:

The correct frequentist interpretation is that a procedure constructed this way captures the fixed parameter in 100(1 − α)% of repeated samples. It is not strictly correct to say that there is a 95% probability that a fixed parameter lies in the already calculated interval.

A prediction interval estimates a range for a future individual observation. For a future observation from a normal population with unknown mean and variance:

The additional term 1 reflects individual future variation. Consequently, prediction intervals are wider than confidence intervals for the mean. In forecasting, intervals should also account for model error, parameter estimation and possible future changes; a textbook interval based only on independent historical observations may be too narrow.

Worked example. Suppose 25 independent daily observations have x̄ = 100 units and s = 20 units. A rough 95% confidence interval for the mean uses t₍.975,24₎ ≈ 2.064:

100 ± 2.064(20/√25) = 100 ± 8.256.

Thus the interval is approximately 91.7 to 108.3 units. The corresponding individual-day prediction interval is:

100 ± 2.064(20√(1+1/25)) ≈ 100 ± 41.6,

or approximately 58.4 to 141.6 units. The mean can be estimated relatively precisely while an individual day remains highly uncertain. This distinction is central to safety-stock decisions.

8. Signal, noise and forecast error

A demand observation can be conceptualised as:

The signal may include level, trend, seasonality, calendar effects, promotions and structural changes. Noise is variation not predictable with the available information. The division is model-dependent: a pattern that appears as noise in a simple model may become signal after incorporating a relevant driver.

Forecast errors are commonly defined as:

where yₜ is actual demand and ŷₜ is the forecast made for period t. The mean error is:

and indicates bias. Mean absolute error is:

Root mean squared error is:

MAE is in demand units and is easier to interpret. RMSE penalises large errors more heavily. Percentage metrics such as MAPE can be unstable or undefined when actual demand is zero or near zero. Alternatives include weighted absolute percentage error, scaled errors and service-level measures.

Residual diagnostics should examine whether errors have non-zero mean, changing variance, autocorrelation or systematic patterns by product, location or time. If residuals are autocorrelated, the model has left predictable structure unexplained. If variance increases with level, a transformation or distributional model may be appropriate. A model that performs well on training data but poorly on future holdout data is overfit.

9. Uncertainty in logistics decisions

Forecast uncertainty has at least three components:

• Aleatory uncertainty: inherent randomness in future demand.
• Epistemic uncertainty: limited knowledge, sparse data or uncertain parameters.
• Structural uncertainty: uncertainty about whether the model form remains valid.

A point forecast is therefore incomplete. A decision-maker may need a quantile, scenario range or service-level probability. For example, if replenishment lead time is L periods and demand during lead time is Dᴸ, a reorder point can be expressed as:

where Qₚ is the pth quantile chosen to meet a target cycle-service probability under the model. Under a simplified normal assumption with mean μᴸ and standard deviation σᴸ:

This formula assumes the lead-time demand distribution is approximately normal and that the chosen quantile corresponds to the desired service probability. It should not be applied mechanically to intermittent demand or when lead time is itself highly variable. If demand and lead time are independent, a common approximation is:

where μᴅ and σ²ᴅ are the per-period demand mean and variance, and E(L) and Var(L) are the mean and variance of lead time. This expression follows from the random-sum variance formula and requires independence; dependence requires a covariance term or simulation.

Uncertainty must be communicated in decision language: “There is a 95% prediction interval of 58–142 units for an individual day” is more useful than “the standard error is 4 units,” provided the assumptions are stated.

10. Illustrative case: regional spare-parts distribution

Background

A regional distributor supplies replacement pumps to industrial customers. Daily demand is usually low, but failures can produce bursts. Management plans to maintain a 95% probability of avoiding a stockout during a five-day replenishment lead time.

Problem

The analyst observes 180 daily shipment records with a mean of 12 units and standard deviation of 9 units. However, the warehouse experienced two stockouts, and shipments may therefore understate true customer demand. Supplier lead time ranges from four to seven days. Management initially proposes a reorder point based on 12 × 5 = 60 units.

Analysis

The proposed point is only the expected demand under a fixed five-day lead time. It ignores demand variability and lead-time variability. It also treats shipments as demand, potentially introducing censoring bias. The analyst first compares customer-order records, lost-sales reports and shipments, then separates emergency projects from routine demand. A time plot reveals no stable seasonal pattern but does show two maintenance-contract events.

Because demand is intermittent and skewed, the analyst does not assume a normal distribution without testing. A bootstrap simulation resamples historically comparable demand days and samples lead times from the observed supplier distribution. It generates the distribution of total demand during each simulated lead time. The reorder point is set to the simulated 95th percentile, subject to review of holding and shortage costs.

Approach and solution

The team adopts three controls:

1. Use censored-demand adjustments and flag stockout periods.
2. Report a median, mean, 95th percentile and prediction range rather than only a mean.
3. Recalculate the reorder point monthly and after contract-driven events.

The inventory decision is not “the forecast says 60.” It is “the selected policy covers approximately 95% of simulated lead-time demand under stated assumptions, and the remaining risk is economically acceptable.”

Outcome

A pilot reduces emergency shipments while increasing average inventory moderately. More importantly, managers can distinguish routine demand risk from exceptional contract events. The policy includes an override process for known maintenance campaigns rather than allowing a single extreme event to permanently inflate the baseline.

Learning points

The case illustrates that data definition precedes calculation; shipment history may not equal unconstrained demand; lead-time uncertainty matters; empirical or simulation-based intervals may be preferable to an unverified normal model; and a service target is a managerial choice, not a purely statistical fact.

11. Limitations, governance and contemporary practice

Statistical methods can create false confidence when data are biased, definitions change or model assumptions are hidden. Common failures include confusing correlation with causation, treating repeated time observations as independent, using random train-test splits for time series, deleting inconvenient outliers, and reporting narrow intervals from a model that excludes structural uncertainty.

Governance requires documented data lineage, versioned forecasts, reproducible code, clear ownership of overrides and monitoring for drift. Forecasts that influence allocation should be audited for systematic bias across regions, customer groups or product classes. Privacy principles require minimisation and appropriate access when order data can identify customers or operational patterns.

AI and machine-learning systems can improve nonlinear prediction, but they do not eliminate uncertainty. Probabilistic forecasting, conformal prediction, quantile models and simulation can produce calibrated ranges. Calibration should be tested: among cases labelled 90% prediction intervals, approximately 90% should contain future observations under stable conditions. Explainability is especially important where managers must justify inventory or service decisions.

Best practice is to use a hierarchy of evidence: establish a transparent baseline, compare alternative models using time-ordered validation, inspect residuals, evaluate both accuracy and calibration, and select the method that improves the decision rather than merely the metric. Forecast intervals should widen when data are sparse, the horizon is long or structural change is plausible.

Chapter Summary

Statistical foundations enable logistics professionals to describe demand, represent uncertainty and make defensible decisions. Means and medians describe centre; variance, standard deviation, percentiles and the coefficient of variation describe dispersion. Probability rules formalise uncertain events, while binomial, Poisson, normal, lognormal and empirical distributions provide alternative representations whose assumptions must be checked. Covariance and correlation describe association but do not establish causality.

Sampling creates estimation error, and the central limit theorem explains when mean-based approximations are useful. Confidence intervals concern population parameters, whereas prediction intervals concern future observations and are wider because they include individual variation. Forecast error analysis separates bias, typical error and the influence of large misses. Signal and noise are model-dependent concepts, and residual diagnostics help identify unexplained structure.

In logistics, uncertainty should be connected to service levels, reorder points, capacity and cost. Sound practice combines quantitative intervals with data governance, scenario analysis, calibrated probabilistic forecasts and managerial judgement.

Key Terms

Population: Complete set of units or outcomes under study.

Sample: Subset of a population used for analysis.

Parameter: Numerical characteristic of a population.

Statistic: Numerical quantity calculated from a sample.

Mean: Arithmetic average of observations.

Median: Middle value of ordered observations.

Variance: Average squared deviation from a mean, with the chosen population or sample denominator.

Standard deviation: Square root of variance, expressed in original units.

Coefficient of variation: Standard deviation divided by a positive mean.

Percentile: Value below which a specified percentage of observations falls.

Random variable: Numerical outcome whose value is uncertain.

Expected value: Probability-weighted long-run average.

Poisson distribution: Count distribution characterised by a rate parameter when its assumptions are reasonable.

Correlation: Standardised measure of linear association.

Regression: Model relating an outcome to one or more predictors.

Standard error: Estimated sampling variability of a statistic.

Confidence interval: Interval produced by a procedure designed to capture a population parameter at a stated long-run rate.

Prediction interval: Interval intended to contain a future individual observation.

Residual: Difference between an observed value and its fitted or forecast value.

Calibration: Agreement between stated probabilistic coverage and observed coverage.

Review Questions

Short Answer Questions

1. Distinguish a parameter from a statistic using a logistics example.
2. Why might the median be preferable to the mean for highly skewed order sizes?
3. State two assumptions of the Poisson distribution.
4. Explain why a prediction interval is wider than a confidence interval for a mean.
5. Why does correlation not establish causation?

Long Answer Questions

1. Explain how descriptive statistics and time plots should be combined when analysing daily demand.
2. Compare the binomial, Poisson and normal distributions, including their logistics applications and limitations.
3. Explain sampling error, standard error and the central limit theorem in the context of demand estimation.
4. Discuss how residual diagnostics can reveal that a forecasting model has failed to capture signal.
5. Develop a governance framework for communicating forecast uncertainty to supply-chain managers.

Analytical / Application Questions

1. A product has daily demand mean 80 units and sample standard deviation 16 units over 36 independent observations. Calculate the approximate standard error of the mean and explain what it measures.
2. A manager proposes setting safety stock equal to two standard deviations of daily demand without considering lead time. Critically evaluate the proposal and identify the additional information required.
3. A forecast has low MAE but persistent positive errors during promotions. Diagnose the likely problem and recommend a validation and modelling response.

Multiple Choice Questions with Answers

Question 1

For observations 8, 10, 10, 12 and 25, which value is the median?

A. 8

B. 10

C. 13

D. 25

Correct Answer: B. 10

Question 2

If a sample has standard deviation 6.86 and mean 13, its coefficient of variation is approximately:

A. 0.053

B. 0.528

C. 1.895

D. 6.86

Correct Answer: B. 0.528

Question 3

Which condition is essential to the standard Poisson model?

A. The variable must be continuous

B. The mean must equal zero

C. Events occur with a stable average rate under the model assumptions

D. All observations must be normally distributed

Correct Answer: C. Events occur with a stable average rate under the model assumptions

Question 4

A correlation of −0.80 indicates:

A. Strong negative linear association

B. Perfect causation

C. No relationship of any kind

D. That both variables have negative means

Correct Answer: A. Strong negative linear association

Question 5

For 25 observations with sample standard deviation 20, the standard error of the mean is:

A. 0.8

B. 4

C. 20

D. 100

Correct Answer: B. 4

Question 6

Which interval is intended to contain a future individual demand observation?

A. Confidence interval for the mean

B. Prediction interval

C. Interquartile range only

D. Correlation interval

Correct Answer: B. Prediction interval

Question 7

The forecast error eₜ = yₜ − ŷₜ is positive when:

A. The forecast exceeds actual demand

B. Actual demand exceeds the forecast

C. The forecast is exactly zero

D. The model has no residuals

Correct Answer: B. Actual demand exceeds the forecast

Question 8

Which metric penalises large forecast errors more heavily because it squares them before averaging?

A. MAE

B. RMSE

C. Median absolute error

D. Mean error sign

Correct Answer: B. RMSE

Question 9

If demand variance is much greater than its mean, a basic Poisson model may suffer from:

A. Underdispersion only

B. Overdispersion

C. Perfect calibration

D. Zero sampling error

Correct Answer: B. Overdispersion

Question 10

A 95% confidence interval should be interpreted as:

A. A 95% probability that the fixed parameter is inside this completed interval

B. A procedure that captures the parameter in about 95% of repeated samples under its assumptions

C. A guarantee that 95% of future observations lie inside it

D. Proof that the data are normally distributed

Correct Answer: B. A procedure that captures the parameter in about 95% of repeated samples under its assumptions