Descriptive, Diagnostic and Process Analytics

Introduction

Operations and supply chains generate large volumes of transactional data: orders, inventory movements, purchase orders, production events, transport milestones, customer complaints and financial postings. Data alone, however, does not explain performance. Analytics creates value when it converts records into evidence about what happened, why it happened and where intervention will produce improvement.

Descriptive, diagnostic and process analytics form an important progression. Descriptive analytics establishes the facts; diagnostic analytics investigates causes and relationships; process analytics examines how work actually flows through an organisation. Together, they support decisions concerning service, cost, speed, quality, capacity, resilience and working capital.

The analyst’s task is not simply to produce attractive dashboards. It is to define the performance question, prepare reliable data, select appropriate methods, interpret variation carefully and translate findings into operational action.

Learning Objectives

After completing this chapter, the learner should be able to:

1. Distinguish descriptive, diagnostic and process analytics.
2. Prepare operational data for exploratory analysis.
3. Select and interpret appropriate summary statistics and visualisations.
4. Segment customers, products, suppliers or processes for managerial action.
5. Apply Pareto analysis to prioritise improvement opportunities.
6. Explain process mining, process discovery and conformance checking.
7. Use bottleneck, variability and root-cause methods responsibly.
8. Evaluate analytical findings in a meaningful supply-chain case.
9. Recognise limitations, governance risks and ethical issues in operational analytics.

1. The Analytical Progression

Descriptive analytics answers: What happened? Examples include monthly order volume, average supplier lead time, stock-out frequency and on-time-in-full delivery.

Diagnostic analytics answers: Why did it happen? It examines associations, differences between groups, sequence patterns, exceptions and plausible causal mechanisms. For example, late deliveries may be associated with a particular carrier, lane, product family or order-release policy.

Process analytics answers: How did work flow through the system? It uses event data to reconstruct actual sequences, waiting times, rework loops, handoffs and deviations from standard procedures.

These forms should not be confused with predictive or prescriptive analytics. A diagnostic association is not automatically a causal explanation, and a process map is not automatically an improvement plan. Strong practice links the analytical result to operational knowledge and a controlled intervention.

A useful analytical cycle is:

1. Define the decision and performance measure.
2. Identify the process, population and time period.
3. Acquire and validate data.
4. Explore distributions, missingness and anomalies.
5. Segment the population and compare performance.
6. Identify concentration, bottlenecks and likely drivers.
7. Validate explanations with process owners.
8. Implement an intervention and monitor the result.

2. Data Foundations for Operational Analysis

Operational data commonly combines master data, transaction data and event data. Master data describes relatively stable entities such as products, locations, suppliers and customers. Transaction data records business events such as sales orders, receipts and shipments. Event data records activities and timestamps, for example “order approved” or “pick completed.”

Before analysis, the analyst must establish:

• Unit of analysis: order, order line, shipment, pallet, production batch or customer.
• Population: the records included and excluded.
• Observation window: dates and time zone.
• Grain: what one row represents.
• Measure definitions: for example, whether an order is on time by requested date or promised date.
• Data lineage: source system, transformation and responsible owner.

A frequent error is mixing grains. If an order contains five lines, joining order-level revenue to line-level records may multiply revenue five times. Grain must therefore be stated before joining tables.

Data-quality checks

Important checks include completeness, validity, uniqueness, consistency and timeliness. Examples are:

• missing carrier codes;
• negative quantities where returns are not expected;
• duplicate shipment identifiers;
• delivery timestamps preceding dispatch timestamps;
• inconsistent units such as kilograms and pounds;
• obsolete product or location codes;
• records posted late or backdated.

Missing values should not automatically be replaced by zero. A missing delivery date, for example, may mean “not yet delivered,” “not captured” or “not applicable.” Each interpretation has a different analytical consequence.

3. Descriptive Statistics for Supply-Chain Data

Let observations be x₁, x₂, …, x_n. The arithmetic mean is:

Here, n is the number of observations and x̄ has the same units as xᵢ. The mean is useful when values are reasonably symmetric, but it is sensitive to extreme observations.

The median is the middle ordered observation, or the average of the two middle observations when n is even. It is often more representative for skewed measures such as delivery lead time or order value.

The range is max(xᵢ) − min(xᵢ). The sample variance and sample standard deviation are:

The denominator n − 1 is used when estimating population variability from a sample. Standard deviation is expressed in the original units; variance is expressed in squared units. These measures assume numerical observations and are most interpretable when the population and sampling process are clearly defined.

Percentiles are particularly useful operationally. The 95th percentile of order-to-delivery time is a value at or below which approximately 95% of observations fall. It does not mean that every order will be delivered within that time. Percentiles are preferable to averages when service commitments depend on tail performance.

For a binary service outcome, such as on-time delivery, the rate is:

where k is the number of successful outcomes and n is the number of eligible cases. The denominator must exclude cases that are genuinely not applicable, and the inclusion rule must be stable over time.

Worked example: lead-time variation

Suppose five supplier lead times are 4, 5, 5, 6 and 10 days. The mean is 6 days and the median is 5 days. The sample variance is:

s² = [(4−6)² + (5−6)² + (5−6)² + (6−6)² + (10−6)²]/4 = 6.5 days².

Thus, s ≈ 2.55 days. The mean is pulled upward by the 10-day observation, while the median describes the typical order more closely. The long tail nevertheless matters for safety stock and customer commitments.

4. Exploratory Analysis and Visualisation

Exploratory data analysis seeks patterns before formal modelling. It should examine distributions, time trends, relationships, outliers and differences among operational segments.

A box plot summarises the median, quartiles and potential outliers. A scatter plot can reveal association, but correlation does not establish causation. For two numerical variables, Pearson correlation is commonly written as:

The coefficient ranges from −1 to +1 and measures linear association. It can be distorted by outliers, nonlinearity, restricted ranges and confounding variables. It should therefore be accompanied by a plot and process reasoning.

Time-series interpretation

A run chart can show trends, cycles, level shifts and unusual points. Analysts should distinguish:

• Common-cause variation: inherent variation from the current system.
• Special-cause variation: variation associated with a specific, identifiable change or event.
• Seasonality: recurring variation linked to calendar or business cycles.
• Trend: sustained directional movement.

A control chart is useful when a stable process and rational sampling method exist. Control limits are not specification limits. Control limits describe expected process variation; specification limits describe customer or engineering requirements. Treating them as interchangeable can result in incorrect decisions.

5. Segmentation and Comparative Diagnosis

Segmentation divides a heterogeneous population into meaningful groups so that averages do not conceal important differences. Common segmentation variables include product family, customer class, supplier, geography, channel, transport mode, order size, demand variability and service requirement.

Good segments are:

• operationally meaningful;
• sufficiently large for reliable comparison;
• mutually exclusive or explicitly overlapping;
• stable enough to support action;
• linked to a decision or intervention.

For example, comparing average supplier lead time alone may be misleading. A supplier with a longer but highly reliable lead time may be preferable to one with a shorter but volatile lead time. Compare central tendency, dispersion, tail percentiles, volume and business criticality together.

A simple customer or product classification may combine annual usage value and demand variability. ABC analysis ranks items by cumulative value, while variability classes may use coefficient of variation:

where s is standard deviation and x̄ is the mean. The coefficient is dimensionless and is meaningful only when the mean is positive and not close to zero. ABC thresholds, such as 70% or 80% of cumulative value, are managerial conventions rather than universal laws.

Comparative analysis should consider exposure. A supplier with 20 late shipments out of 100 is not equivalent to one with 20 late shipments out of 1,000. Rates, confidence intervals or suitable statistical tests may be needed when observed differences are close or sample sizes differ greatly.

6. Pareto Analysis

Pareto analysis prioritises categories by arranging them in descending order of impact and calculating cumulative contribution. It is based on the practical observation that effects are often concentrated, not on a universal requirement that exactly 80% of consequences arise from 20% of causes.

Worked example

A distributor records annual stock-out incidents by cause:

The first three categories account for 85% of incidents. Improvement work should initially examine these causes, but the result is not proof that they are root causes. Cause labels may be inconsistent, incidents may have multiple causes, and frequency may not reflect financial or customer impact. A second Pareto analysis by lost margin, expedited freight or criticality may produce a different priority.

7. Process Analytics and Process Mining

Process analytics studies the sequence, timing and outcomes of activities. Process mining uses event logs to discover or evaluate processes from actual system data.

A minimally useful event log contains:

1. Case identifier: the unit moving through the process, such as order ID.
2. Activity: event name, such as approve, pick or ship.
3. Timestamp: when the activity occurred.
4. Optional attributes: user, location, product, value, channel or outcome.

A process-mining event log can reveal the difference between the documented process and the process actually followed. Common findings include repeated approvals, skipped controls, rework loops, long queues, parallel activities and process variants by site or customer.

Three related uses are:

• Process discovery: infer a process model from event data.
• Conformance checking: compare observed traces with a reference model or rule set.
• Enhancement: enrich a process model with timing, cost, frequency or outcome information.

A process map should distinguish activity time from waiting time. Total cycle time can be represented as:

where all terms use the same time unit and cover the same case. In many administrative and supply-chain processes, waiting time is larger than touch time. Reducing touch time alone may therefore have little effect on customer lead time.

A process variant is a distinct sequence of activities. Too many variants may indicate unnecessary customisation, inconsistent policy or poor master-data discipline. However, not every variant is wasteful; urgent, regulated or high-value orders may legitimately require different treatment.

Bottlenecks and flow

A bottleneck is a resource or policy constraint that limits system throughput. It may be a machine, labour pool, approval queue, dock, supplier, data interface or planning rule. The slowest activity is not always the bottleneck: a resource’s effective capacity, demand load, variability and interaction with downstream constraints must be examined.

For a stable production or service flow, Little’s Law is:

where L is average work-in-process, λ is average throughput rate and W is average time in the system. Units must be consistent, such as orders, orders per day and days. The relationship assumes a stable system over the period analysed, with average inflow and outflow balanced and a clearly defined system boundary. It is not a short-term guarantee for an unstable or rapidly changing system.

For example, if an average of 120 orders are in a fulfilment process and throughput is 30 orders per day, average cycle time is W = L/λ = 120/30 = 4 days. Reducing inventory in process without increasing throughput does not automatically reduce cycle time; the relationship indicates that at least one of the variables must change.

8. Root-Cause Analysis

Root-cause analysis seeks factors whose removal or control prevents recurrence. It should not stop at the first plausible explanation. Useful methods include:

Five Whys

The analyst repeatedly asks why an event occurred. The method is conversational and useful for simple causal chains, but it can become subjective if several causes exist or if the team forces a single linear chain.

Fishbone diagram

Causes are organised under categories such as people, process, technology, materials, measurement and environment. The categories are prompts, not proof. Each proposed cause requires evidence.

Fault-tree thinking

The analyst begins with an undesired event and decomposes it through logical AND and OR relationships. This is valuable for complex failures because multiple conditions may be necessary.

Evidence matrix

A practical diagnostic table records each hypothesis, supporting evidence, contradictory evidence, data required and proposed test. This discourages confirmation bias.

A disciplined investigation distinguishes:

• symptom: what is observed;
• proximate cause: the immediate mechanism;
• systemic cause: the condition allowing recurrence;
• control failure: why the issue was not detected or prevented.

For example, “late shipment” is a symptom. “Carrier missed collection” may be proximate. “Orders are released after the carrier cut-off because the planning system uses local time incorrectly” is closer to a systemic explanation.

9. Integrated Applied Case: Northstar Distribution

This is an illustrative case. Northstar Distribution supplies maintenance parts to industrial customers. Its on-time shipment rate fell from 94% to 86% over one quarter, while expedited freight cost increased by 28%.

Background and problem

Managers initially suspected warehouse labour shortages. The available data contained order ID, customer, warehouse, order-release timestamp, pick completion, carrier collection, promised date and actual dispatch. The team first defined on-time shipment as dispatch on or before the promised date and excluded cancelled orders.

Analysis

Data-quality checks found duplicate shipment records and 7% of orders with missing carrier collection times. The duplicates were removed, while missing collection times were retained as a separate data-quality category rather than treated as zero duration.

Descriptive analysis showed that the mean order cycle time increased from 1.8 to 2.4 days, but the 95th percentile increased from 3.6 to 6.9 days. The deterioration was therefore concentrated in the tail. A Pareto analysis of late orders found that 61% were associated with three causes: late order release, inventory-record discrepancy and carrier cut-off miss.

Segmentation showed that Warehouse B accounted for 52% of late orders while processing 31% of volume. The problem was most severe for small, urgent orders released after 15:00. A process-mining analysis identified a frequent loop: pick started, inventory discrepancy recorded, replenishment requested, pick restarted. The median waiting time in this loop was 19 hours.

Approach and solution

The team implemented four actions:

1. Corrected the local-time configuration used in order release.
2. Introduced a same-day cycle count for high-frequency discrepancy locations.
3. Created an exception queue for urgent orders approaching carrier cut-off.
4. Required a reason code for manual promise-date changes.

The team did not immediately add labour because the evidence indicated that waiting and data accuracy, rather than total picking capacity, were dominant drivers. A two-week pilot at Warehouse B was compared with the preceding baseline and a similar warehouse that had not yet changed its process.

Outcome

After four weeks, Warehouse B’s on-time shipment rate rose to 93%, the 95th-percentile cycle time fell to 4.1 days and expedited freight cost declined by 16%. The comparison warehouse showed no similar improvement. Because the pilot was not a fully randomised experiment and other changes occurred during the period, the result supports—but does not conclusively prove—the intervention’s causal effect. Northstar therefore continued monitoring and audited the data definitions monthly.

Learning points

The case illustrates five principles: define the metric before analysing; inspect the tail rather than only the mean; segment performance by operational context; use event sequences to reveal waiting and rework; and validate analytical claims through controlled comparison and process-owner review.

10. Technology, Governance and Ethics

Modern analytics platforms combine SQL, business-intelligence tools, process-mining software, cloud data warehouses and machine-learning assistants. AI can help classify text complaints, identify unusual patterns or generate candidate explanations. It should not be allowed to present unverified correlations as causes or to change operational policy without accountable human review.

Key governance practices include:

• documented metric definitions and data lineage;
• role-based access to employee, customer and supplier data;
• minimisation of personally identifiable information;
• audit trails for transformations and dashboard changes;
• validation of automated classifications;
• retention and deletion rules;
• monitoring for biased treatment of suppliers, workers or customers;
• clear ownership of decisions based on analytics.

Employee-level process data can create surveillance risks. A dashboard that ranks individuals without accounting for task complexity, equipment availability or assignment patterns may be unfair and operationally harmful. Analytics should focus on improving systems where possible, use appropriate aggregation, and communicate purpose and safeguards.

11. Limitations and Best Practices

Descriptive and diagnostic analysis face several limitations. Data may reflect only recorded work, not informal work. Averages can hide tails and subgroups. Historical associations may change after a policy intervention. Survivorship bias can exclude failed suppliers or cancelled orders. Simpson’s paradox can occur when an aggregate relationship reverses after segmentation. Confounding variables may create a misleading association.

Best practices are therefore to:

1. Start with a decision, not a chart.
2. State the unit, grain, population and time window.
3. Examine distributions and missingness before aggregation.
4. Report volume, rate and exposure together.
5. Compare central tendency with tail performance.
6. Segment using operationally meaningful dimensions.
7. Treat Pareto categories as priorities for investigation, not proven causes.
8. Separate waiting, processing, transport and rework time.
9. Validate findings with frontline expertise and independent evidence.
10. Use pilots, control groups or before-and-after designs cautiously.
11. Preserve reproducibility through versioned data and documented logic.
12. Monitor whether improvement persists and whether costs or risks shift elsewhere.

Chapter Summary

Descriptive analytics establishes what happened through reliable definitions, summary statistics and visualisation. Diagnostic analytics investigates why performance differs by examining segments, relationships, concentration and plausible causal mechanisms. Process analytics adds the dimension of sequence and time, revealing actual process variants, queues, rework and conformance gaps.

Operational analysis depends on data grain, quality, exposure and context. Means should be complemented by medians, percentiles and variability measures. Pareto analysis helps prioritise concentrated effects but does not establish root cause. Segmentation makes heterogeneous performance visible, while process mining reconstructs activities from case-based event logs. Little’s Law links work-in-process, throughput and cycle time under stable conditions. Root-cause methods are most effective when hypotheses are tested against evidence rather than accepted from intuition.

The professional analyst combines technical methods with process knowledge, governance, ethical judgment and controlled learning. The ultimate purpose is not explanation alone, but better decisions and measurable operational improvement.

Key Terms

Descriptive analytics: Analysis that summarises what has happened.

Diagnostic analytics: Analysis that investigates reasons and drivers of observed performance.

Process analytics: Analysis of activities, sequences, timing and outcomes in a process.

Event log: Records containing a case, activity and timestamp, often with additional attributes.

Grain: The level of detail represented by one record.

Unit of analysis: The entity about which an observation or conclusion is made.

Mean: Arithmetic average of numerical observations.

Median: Middle value in an ordered set.

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

Standard deviation: Measure of dispersion in the original measurement units.

Coefficient of variation: Standard deviation divided by the mean.

Segmentation: Division of a population into meaningful comparison groups.

Pareto analysis: Ranking categories by impact and displaying cumulative contribution.

Process discovery: Inferring a process model from event data.

Conformance checking: Comparing observed process behaviour with a reference model or rule.

Process variant: A distinct sequence of activities in a process.

Bottleneck: Constraint that limits system throughput.

Root cause: Underlying condition whose correction prevents recurrence.

Common-cause variation: Variation inherent in the current system.

Special-cause variation: Variation attributable to an identifiable unusual factor.

Review Questions

Short Answer Questions

1. What are the principal questions answered by descriptive and diagnostic analytics?
2. Why must the grain of a dataset be established before joining tables?
3. Distinguish specification limits from control limits.
4. What three fields are minimally required in a process-mining event log?
5. Why should a Pareto result not automatically be treated as a root-cause analysis?

Long Answer Questions

1. Explain how mean, median, standard deviation and percentiles support analysis of supplier lead time.
2. Discuss the design principles of effective operational segmentation and the risks of poorly chosen segments.
3. Explain process discovery, conformance checking and process enhancement, giving a supply-chain example for each.
4. Derive the managerial meaning of L = λW and state the assumptions required for its use.
5. Compare Five Whys, fishbone analysis, fault-tree thinking and an evidence matrix as root-cause methods.

Analytical / Application Questions

1. A warehouse has 200 orders in process and completes 50 orders per day. Estimate average cycle time using Little’s Law. State one condition under which the estimate may be misleading.
2. A supplier has an average lead time of 8 days and a standard deviation of 4 days; another has an average of 12 days and a standard deviation of 3 days. Calculate each coefficient of variation and recommend what additional evidence is needed before selecting a supplier.
3. A dashboard shows that one warehouse has the worst on-time rate. Design an analytical investigation that tests whether the result is caused by order mix, promise-date policy, inventory accuracy, labour capacity or carrier performance.

Multiple Choice Questions with Answers

Question 1

Which question is primarily answered by descriptive analytics?

A. What happened?

B. What intervention will optimise the network?

C. What will demand be next year?

D. Which causal model is statistically true?

Correct Answer: A. What happened?

Question 2

In an operational event log, what identifies the process instance to which an activity belongs?

A. Case identifier

B. Specification limit

C. Control limit

D. Cumulative percentage

Correct Answer: A. Case identifier

Question 3

For observations 4, 5, 5, 6 and 10 days, which measure is least affected by the extreme value 10?

A. Mean

B. Median

C. Range

D. Variance

Correct Answer: B. Median

Question 4

A Pareto chart is principally used to:

A. Prove a causal relationship

B. Prioritise categories by contribution

C. Calculate process capacity

D. Replace data-quality checks

Correct Answer: B. Prioritise categories by contribution

Question 5

Which statement about correlation is correct?

A. Correlation always proves causation

B. Correlation measures only whether data are complete

C. Correlation can show association but does not by itself establish causation

D. Correlation is valid only for categorical variables

Correct Answer: C. Correlation can show association but does not by itself establish causation

Question 6

Which activity is an example of process conformance checking?

A. Inferring a process model from event data

B. Comparing observed sequences with an approved process model

C. Calculating a product’s annual demand value

D. Replacing missing values with zero

Correct Answer: B. Comparing observed sequences with an approved process model

Question 7

If average work-in-process is 120 orders and throughput is 30 orders per day, Little’s Law gives an average cycle time of:

A. 2 days

B. 3 days

C. 4 days

D. 6 days

Correct Answer: C. 4 days

Question 8

Which is most likely a systemic root cause rather than a symptom?

A. A customer received an order late

B. A picker made one error

C. The order-release system uses the wrong local time zone

D. A complaint was recorded

Correct Answer: C. The order-release system uses the wrong local time zone

Question 9

A coefficient of variation is calculated as:

A. Mean divided by standard deviation

B. Standard deviation divided by mean

C. Range divided by sample size

D. Median divided by percentile

Correct Answer: B. Standard deviation divided by mean

Question 10

Which practice best supports responsible operational analytics?

A. Treat every missing value as zero

B. Publish employee rankings without context

C. Document metric definitions, data lineage and access controls

D. Use the average alone to represent process performance

Correct Answer: C. Document metric definitions, data lineage and access controls