Demand, Performance and Decision Analytics

Introduction

Product and portfolio management increasingly depends on the ability to convert imperfect market evidence into disciplined decisions. Brand managers must estimate the size of an opportunity, forecast demand, determine whether customers are adopting and remaining loyal, evaluate satisfaction, monitor market share, and decide where scarce investment should be allocated. These tasks are connected, but they are not interchangeable. A large market is not necessarily attractive; high usage is not necessarily profitable; correlation between a campaign and sales is not proof that the campaign caused the sales increase.

Decision analytics provides the concepts, metrics, methods, and governance needed to address these issues. It combines market research, behavioral data, statistical modelling, experimentation, financial analysis, and managerial judgment. The objective is not to produce a dashboard full of numbers. It is to improve the quality, speed, transparency, and reversibility of decisions while communicating uncertainty honestly.

At graduate level, analytics should be treated as both a technical and organizational capability. A technically sophisticated model can fail if its data-generating process is misunderstood, if incentives encourage metric manipulation, or if managers confuse prediction with explanation. This chapter therefore considers measurement design, causal reasoning, portfolio interpretation, and ethical use of data together.

Learning Objectives

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

1. Distinguish market potential, addressable market, demand, sales, and share.
2. Construct defensible bottom-up and top-down market-size estimates.
3. Select and interpret forecasting methods under uncertainty.
4. Use cohorts, funnels, retention, satisfaction, and portfolio metrics appropriately.
5. Design and evaluate experiments, distinguishing correlation from causation.
6. Build decision-oriented dashboards rather than descriptive scoreboards.
7. Communicate confidence intervals, scenarios, assumptions, and model risk.
8. Integrate evidence into product, brand, and portfolio decisions.

1. The Architecture of Decision Analytics

A useful analytical architecture follows the decision chain:

Market opportunity → customer acquisition → activation and usage → retention → value and profitability → portfolio allocation.

Each stage requires a different question:

The architecture is not a claim that all products follow a linear path. Industrial products may have long procurement cycles, luxury products may have infrequent purchase, and services may generate value through continuous use. The framework is a starting point for mapping the actual customer and economic process.

Measurement design

A metric is useful only when its definition, population, time window, unit, and decision use are explicit. For example, “monthly active user” should specify the qualifying activity, the calendar convention, treatment of duplicate accounts, and whether employees or test accounts are excluded. A metric specification should include:

• Name and purpose: what decision it informs.
• Numerator and denominator: especially for rates.
• Population: customers, prospects, accounts, households, or transactions.
• Time period: calendar month, rolling 30 days, cohort age, or fiscal quarter.
• Data source and refresh rate.
• Owner and quality checks.
• Known limitations and exclusions.

This discipline prevents metric drift, in which a measure appears to improve because its definition changed.

2. Market Sizing and Demand Estimation

Market sizing estimates the economic or behavioral scope of an opportunity. Four concepts must be separated:

• Potential market: all possible users or occasions, including those not currently served.
• Total addressable market (TAM): the value or volume if the relevant offering captured the entire defined market.
• Serviceable available market (SAM): the portion reachable through geography, capabilities, regulation, price, or channel.
• Serviceable obtainable market (SOM): the realistically attainable portion within a specified period and competitive context.

These are scenarios, not guaranteed outcomes. A market-size estimate becomes misleading when an unconstrained TAM is presented as a near-term sales forecast.

Bottom-up sizing

A bottom-up estimate builds from observable units. A general value estimate is:

where N is the number of relevant customers or households, p is the proportion with the relevant need, f is purchase occasions per period, and v is average value per occasion. Units must be consistent: if f is annual frequency and v is dollars per purchase, the result is annual dollars.

Suppose a region contains 2,000,000 households. Research suggests that 30% are in the relevant category, each qualifying household buys an average of 4 units per year, and average revenue per unit is $18. The estimated annual value is:

2,000,000 × 0.30 × 4 × $18 = $43,200,000.

If a new brand expects to reach 5% of that market in three years, the implied revenue scenario is $2.16 million, before considering distribution constraints, cannibalization, competitor response, and price changes. The estimate should be reported with assumptions, not as a precise fact.

Top-down sizing and triangulation

Top-down estimates begin with an industry total and apply relevant filters. They are efficient but may embed inconsistent definitions or optimistic analyst assumptions. Strong practice triangulates at least two approaches:

1. Top-down: published category value × relevant segment proportion.
2. Bottom-up: customers × incidence × frequency × value.
3. Comparable analysis: analogous launches, regions, or categories.
4. Behavioral validation: search, transaction, distribution, or usage evidence.

Differences between methods are diagnostically useful. They may reveal different category boundaries, double counting, unobserved informal consumption, or an incorrect assumption about frequency.

Market share

Value share and volume share answer different questions:

A brand may have higher value share than volume share because it commands a premium price. Share of preference, share of search, and share of voice are related indicators but are not substitutes for realized market share.

3. Forecasting Demand Under Uncertainty

Forecasting is a structured estimate of future outcomes conditional on assumptions. It is not a prediction of what must happen. Forecasts support capacity, inventory, pricing, media, hiring, and portfolio decisions.

Basic forecasting methods

A naïve forecast uses the most recent observation as the next forecast. It is a useful benchmark, not necessarily a sophisticated method. A moving average smooths short-term variation:

where F is the forecast, D is observed demand, t is the current period, and k is the number of periods included. The method assumes that recent observations are informative and that a simple average is appropriate. It can lag when demand is trending.

An exponential smoothing forecast gives greater weight to recent demand:

where 0 ≤ α ≤ 1 is the smoothing parameter. Higher α responds more rapidly to new information but may react excessively to noise. These methods are appropriate for time series with relatively stable structures; they do not automatically account for promotions, stockouts, competitor actions, price changes, or structural breaks.

Forecast accuracy can be summarized by mean absolute error:

where n is the number of evaluated periods. MAE is expressed in the units of demand. Mean absolute percentage error can be difficult when actual demand is zero or close to zero; therefore, no single accuracy metric should be used mechanically.

Causal and judgmental inputs

Historical data may not represent the future. A product launch, regulatory change, recession, competitor withdrawal, or major distribution gain can invalidate a purely extrapolative forecast. Managers should combine statistical baselines with causal drivers and scenario analysis.

A useful forecast has:

• a baseline case;
• upside and downside cases;
• explicit assumptions;
• a prediction interval or probability range where possible;
• trigger conditions for revising the forecast;
• an owner responsible for monitoring error.

Scenario analysis is not the same as assigning arbitrary optimistic and pessimistic numbers. Scenarios should correspond to identifiable states, such as distribution coverage, adoption rate, price realization, or competitor response.

4. Funnel, Cohort, and Customer-Lifecycle Analysis

A funnel describes transitions between stages, while a cohort groups customers according to a shared starting event, such as acquisition month, first purchase, plan type, or geography. Cohort analysis prevents the blending of new and mature customers, which can conceal deterioration or falsely suggest improvement.

Conversion and retention

For a stage transition:

The denominator must be the population eligible to advance, not necessarily all visitors. A retention rate for a cohort can be expressed as:

If 1,000 customers first purchase in January and 620 purchase again in March, two-month repeat retention is 62%, provided the time definition and eligibility rules are clear. Retention may be measured by account activity, subscription status, repeat purchase, or revenue; these measures can produce different conclusions.

Churn is the complement of retention only when the populations and definitions match. Subscription churn may mean cancellation during a period, while behavioral churn may mean no activity for a specified inactivity window. Infrequent categories require longer observation windows and should not be judged by subscription-style definitions.

Cohort interpretation

A cohort table can reveal:

• whether newer acquisition sources retain better or worse;
• whether onboarding changes improve early activation;
• whether retention stabilizes or continues declining;
• whether growth is being purchased through low-quality customers;
• whether price, channel, or product variants affect lifetime value.

Cohort analysis remains observational unless assignment to an intervention is controlled. A cohort exposed to a new campaign may differ systematically from earlier cohorts.

Customer lifetime value

A simplified contribution-based lifetime value for a subscription relationship is:

where m is periodic contribution margin per customer, r is the probability of retention from one period to the next, i is the periodic discount rate, and CAC is customer acquisition cost. This geometric model assumes constant margin, constant retention probability, and a stable discount rate. It is inappropriate when retention changes sharply with tenure, margins vary by customer, or acquisition costs are allocated inconsistently. The result is a decision estimate, not an accounting valuation.

5. Product and Brand Performance Metrics

Metrics should be organized into a hierarchy rather than treated as an undifferentiated list.

Leading and lagging indicators

Leading indicators provide early signals: qualified reach, trial, activation, feature adoption, distribution, search intent, and service response time. Lagging indicators record outcomes: revenue, contribution margin, market share, renewal, recommendation, and brand consideration. Leading indicators are useful only if they have a demonstrated relationship with later outcomes; otherwise they become vanity metrics.

Adoption and engagement

Adoption asks whether a target population begins using a product. Engagement asks how and how deeply it is used. Useful measures include activation rate, frequency, breadth of feature use, usage intensity, repeat purchase, and time to value. High engagement can be harmful if it reflects customer difficulty, excessive service contacts, or forced usage rather than customer benefit.

Satisfaction and loyalty

Customer satisfaction surveys should define the customer experience being evaluated and use reliable sampling procedures. A recommendation score is often computed as:

The score ranges from −100 to +100 when respondents are classified according to the specified response scale. It is not a direct measure of profit, retention, or causal advocacy. Sampling bias, question wording, response timing, cultural effects, and nonresponse can materially affect it. Combining satisfaction with behavioral outcomes and verbatim feedback is more informative than relying on one score.

Portfolio health

A portfolio view should evaluate products across financial, market, customer, and strategic dimensions. A compact scorecard might include:

Portfolio metrics should not be aggregated into a single score unless weights are explicit, validated, and periodically reviewed. A product with low current profit may be strategically important, while a high-share product may be declining and consuming disproportionate resources.

6. Experimentation and Causal Inference

Managers commonly observe that two variables move together and infer causation. This is dangerous because of confounding, reverse causality, selection bias, seasonality, and common external shocks.

• Correlation describes statistical association.
• Causation concerns the change in an outcome attributable to an intervention, relative to a credible counterfactual.

For example, customers who use a premium feature may retain longer. This does not prove the feature causes retention: highly motivated customers may both use the feature and remain loyal.

Randomized experiments

In an A/B test, eligible units are randomly assigned to treatment and control. Randomization aims to balance observed and unobserved characteristics in expectation. The estimated intention-to-treat effect is:

where Y is the outcome and T indicates assignment to treatment (1) or control (0). The estimate is causal under appropriate randomization, implementation, measurement, and interference assumptions. The unit may be a user, account, store, region, or time period; choosing the wrong unit can create contamination or dependence.

A proportion difference can be accompanied by a standard error. For independent treatment and control groups with observed proportions p₁ and p₀ and sample sizes n₁ and n₀, an approximate standard error is:

This approximation assumes sufficiently large samples and independent observations. A confidence interval is an interval procedure under repeated-sampling assumptions; it is not a probability statement that a fixed parameter lies in the particular interval.

Experiment quality

Before launch, specify the primary outcome, eligible population, assignment unit, duration, minimum detectable effect, decision threshold, and stopping rules. Guardrail metrics should detect harms such as complaints, refunds, latency, or unsubscribes. Multiple testing, repeated peeking, noncompliance, novelty effects, and spillovers can inflate false discoveries or limit generalizability.

When randomization is impossible, analysts may use difference-in-differences, interrupted time series, matching, instrumental variables, or regression adjustment. These designs depend on stronger assumptions. For example, difference-in-differences requires a credible parallel-trends assumption: absent treatment, treated and comparison units would have followed similar outcome trends. The assumption should be examined rather than asserted.

7. Dashboards and Decision Communication

A dashboard is an interface for action, not a warehouse for every available measure. An effective dashboard connects:

1. a strategic question;
2. a small set of diagnostic and outcome metrics;
3. targets or comparison groups;
4. trend and cohort views;
5. uncertainty and data-quality indicators;
6. an owner and decision rule.

Dashboards should distinguish actuals, estimates, and forecasts visually and verbally. A red indicator should mean that a predefined threshold has been breached, not merely that a number is lower than last month. Managers should be able to drill from portfolio performance to segment, channel, cohort, and customer experience without losing metric definitions.

Communicating uncertainty requires more than adding decimals. Analysts should state the source of uncertainty: sampling error, model error, demand variability, missing data, measurement error, or scenario dependence. Ranges, fan charts, probability-weighted scenarios, and sensitivity analysis are useful when their construction is explained.

A decision memo can use the following structure:

• Decision required: what must be decided and by when.
• Evidence: key observed facts and definitions.
• Interpretation: what the evidence supports and does not support.
• Alternatives: options, costs, benefits, and risks.
• Uncertainty: assumptions, interval estimates, and failure modes.
• Recommendation: selected action and rationale.
• Learning plan: leading indicators, review date, and reversibility.

8. Applied Case: Aurora Home’s Smart Air Monitor

Illustrative case. Aurora Home, a mid-sized consumer brand, is considering a smart indoor-air monitor priced at $120. Management initially presents a $1.2 billion TAM based on all urban households and proposes a national launch.

Background and problem

The product team must determine whether to launch nationally, test selected regions, or discontinue the concept. Available evidence includes a category report, a six-month prototype test, retailer interviews, and a digital acquisition campaign. The prototype generated strong interest but weak repeat engagement.

Analysis

A bottom-up estimate identifies 8 million reachable households in the initial countries. Research estimates that 25% have a relevant concern, and 40% of those are willing to consider a monitor within two years. The resulting potential buyers are:

8,000,000 × 0.25 × 0.40 = 800,000.

At a $120 list price, the unconstrained two-year revenue opportunity is $96 million, not $1.2 billion. This is still not a forecast: channel access, price discounts, replacement, and competitor response remain uncertain.

Prototype cohorts show that 42% complete setup, but only 18% are active after 90 days. Interviews indicate that the product provides an initial “air quality” score but does not make recommended actions sufficiently clear. Acquisition from a discount-led campaign is cheaper, but those customers retain less well than customers acquired through specialist retailers.

The team proposes a randomized onboarding experiment among new customers. The treatment receives action-oriented explanations and reminders; the control receives the existing interface. The primary outcome is 90-day active use; guardrails are returns, complaints, and notification opt-outs. The test is powered before launch and assigned at the customer-account level to avoid household-level duplication where possible.

Approach and solution

The portfolio committee rejects the immediate national launch and approves a staged regional test. The decision uses three gates:

1. Setup completion must exceed 55%.
2. Ninety-day active use must improve materially without increasing returns.
3. Contribution economics must remain positive after retailer margin, support costs, and acquisition expense.

The forecast is reported as base, upside, and downside scenarios. The dashboard displays acquisition source, cohort retention, setup funnel, product defects, contribution margin, and customer sentiment. The committee also requires a data-retention policy because the device can infer household routines and potentially sensitive environmental conditions.

Outcome

In the illustrative follow-up, treatment increases setup completion to 58% and 90-day activity to 27%, while returns remain statistically indistinguishable from control. However, retailer economics are weaker than direct-to-consumer economics, and the service cost of replacement sensors is higher than expected. Management launches only in two specialist channels, redesigns the replacement program, and postpones mass retail expansion.

Learning points

The case demonstrates that market size does not establish launch readiness. Cohort evidence identified an activation problem, experimentation provided stronger evidence about the onboarding intervention, and portfolio analysis incorporated channel economics and data risk. The final decision was not “launch” or “do not launch” in the abstract; it was a staged commitment with measurable gates and learning objectives.

9. Governance, Ethics, and Contemporary Developments

Analytics can create harm when data are collected without meaningful consent, when sensitive attributes are inferred, or when automated decisions disadvantage groups. Governance should address purpose limitation, data minimization, access control, retention, security, auditability, and rights available under applicable privacy law. Analysts should assess whether a metric encourages surveillance, manipulative design, discriminatory targeting, or exclusion of vulnerable customers.

Algorithmic systems require attention to data leakage, label bias, proxy discrimination, model drift, and explainability. Generative AI can summarize feedback, classify complaints, and assist scenario generation, but generated outputs require human validation. AI should not be treated as an independent source of market truth. Synthetic or scraped data may distort language, geography, or customer representation.

Contemporary practice increasingly combines event-level behavioral data, experimentation platforms, customer-data platforms, causal inference, and machine-learning forecasting. These tools improve speed but do not remove the need for conceptual definitions and identification assumptions. Privacy-enhancing techniques, federated analysis, differential privacy, and stronger consent management are becoming more important as third-party identifiers decline.

Best-practice principles

1. Start with a decision, not with available data.
2. Define metrics before observing results.
3. Separate descriptive, predictive, and causal claims.
4. Use cohorts and segments appropriate to the product’s buying cycle.
5. Compare forecasts with simple benchmarks.
6. Report ranges, assumptions, and model limitations.
7. Predefine experiment outcomes and guardrails.
8. Combine financial, behavioral, customer, and strategic evidence.
9. Audit data quality, privacy, fairness, and access.
10. Use staged investment when uncertainty is material and learning is feasible.

Chapter Summary

Demand, performance, and decision analytics connect market opportunity to portfolio action. Market sizing requires clear boundaries and triangulation; TAM, SAM, and SOM are scenario concepts rather than guaranteed sales. Forecasts should be benchmarked, assumption-driven, and accompanied by uncertainty. Funnel and cohort analysis reveal where adoption and retention change, while customer lifetime value requires explicit assumptions about margin, retention, discounting, and acquisition cost.

Performance metrics should form a hierarchy of leading and lagging indicators and should be interpreted in context. Market share, satisfaction, engagement, and profitability answer different questions. Experiments support causal claims when assignment, measurement, and implementation are credible; observational association does not establish causation. Dashboards are most valuable when they link metrics to decisions, thresholds, owners, and learning plans. Finally, ethical governance, privacy, fairness, and transparent communication are central to responsible analytics. The strongest portfolio decisions combine quantitative evidence with managerial judgment, staged commitments, and explicit recognition of uncertainty.

Key Terms

Market potential: The broadest estimated scope of possible customers or occasions.

TAM: Total addressable market under an unconstrained definition.

SAM: Serviceable available market reachable under specified constraints.

SOM: Serviceable obtainable market realistically attainable in a period.

Market share: A brand’s proportion of category value or volume.

Forecast: A conditional estimate of a future outcome.

Cohort: Customers sharing a defined starting event or characteristic.

Retention: The proportion of an original population remaining active or purchasing at a specified time.

Churn: Loss of customers, subscriptions, or activity under a defined rule.

Activation: Initial achievement of a meaningful product-use milestone.

Engagement: Frequency, depth, or breadth of product interaction.

Customer lifetime value: Estimated future contribution from a customer relationship.

Customer acquisition cost: Cost incurred to acquire a customer under a stated allocation rule.

Leading indicator: An early measure expected to signal a later outcome.

Lagging indicator: A measure of an outcome that has already occurred.

Correlation: Statistical association between variables.

Causation: A change in an outcome attributable to an intervention relative to a counterfactual.

A/B test: A randomized comparison of treatment and control conditions.

Guardrail metric: A measure used to detect unintended harm or deterioration.

Model risk: The possibility that assumptions, data, or implementation make analytical outputs misleading.

Review Questions

Short Answer Questions

1. Distinguish TAM, SAM, and SOM.
2. Why can value share differ from volume share?
3. What is the purpose of cohort analysis?
4. Why does correlation not establish causation?
5. Identify two components that should appear in an uncertainty statement.

Long Answer Questions

1. Explain how a brand manager could triangulate a bottom-up market-size estimate with top-down and comparable-market evidence.
2. Compare moving-average and exponential-smoothing forecasts, including their assumptions and limitations.
3. Discuss how adoption, activation, engagement, retention, satisfaction, and profitability should be connected in a portfolio dashboard.
4. Explain the assumptions and governance requirements of a randomized A/B test.
5. Evaluate the ethical and managerial risks of using AI-generated customer and market analytics.

Analytical / Application Questions

1. A category contains 3,000,000 households. Twenty percent have the relevant need, 50% are reachable through available channels, and expected annual spending is $80 per qualifying household. Estimate the reachable annual market value and identify two reasons it may overstate obtainable revenue.
2. A new cohort contains 4,000 customers, of whom 1,100 are active after six months. Calculate six-month retention and explain why the result should be compared with other cohorts rather than interpreted in isolation.
3. A campaign is associated with higher retention among exposed customers. Design an evaluation that could distinguish campaign causation from self-selection, specifying the assignment unit, primary outcome, guardrail metric, and major threat to validity.

Multiple Choice Questions with Answers

Question 1

Which estimate is most appropriately described as SOM?

A. All households that could theoretically need the category

B. The entire global category value reported by an analyst

C. The portion the firm can realistically obtain within a defined period

D. The number of people who have ever heard of the brand

Correct Answer: C. The portion the firm can realistically obtain within a defined period

Question 2

A market has 2,000,000 households, 30% relevant incidence, four annual purchases, and $18 per purchase. What is the annual market value?

A. $10.8 million

B. $21.6 million

C. $43.2 million

D. $72.0 million

Correct Answer: C. $43.2 million

Question 3

What is the principal analytical purpose of cohort analysis?

A. To eliminate all sampling error

B. To separate customers by a shared starting point and observe outcomes over time

C. To prove that a campaign caused retention

D. To replace financial analysis with behavioral data

Correct Answer: B. To separate customers by a shared starting point and observe outcomes over time

Question 4

Which statement about a moving-average forecast is correct?

A. It gives every included observation equal weight

B. It always responds faster than exponential smoothing

C. It requires randomized treatment assignment

D. It is unaffected by the choice of the number of periods

Correct Answer: A. It gives every included observation equal weight

Question 5

A brand has revenue of $12 million in a $60 million category. What is its value share?

A. 5%

B. 12%

C. 20%

D. 48%

Correct Answer: C. 20%

Question 6

Which condition most directly supports a causal interpretation of an A/B test?

A. Treatment customers have higher average revenue before the test

B. Eligible units are randomly assigned and outcomes are measured as specified

C. The treatment group is larger than the control group

D. The treatment result is consistent with management intuition

Correct Answer: B. Eligible units are randomly assigned and outcomes are measured as specified

Question 7

If 1,000 customers enter a cohort and 620 remain active at the defined measurement point, what is retention?

A. 38%

B. 60%

C. 62%

D. 1,620%

Correct Answer: C. 62%

Question 8

Which is a leading indicator rather than a lagging outcome?

A. Prior-year contribution margin

B. Current market share

C. Completed onboarding rate

D. Recognized quarterly revenue

Correct Answer: C. Completed onboarding rate

Question 9

Why can a high customer lifetime value estimate be misleading?

A. Lifetime value never uses assumptions

B. It may assume constant retention and margin when customer economics change over time

C. It is measured only in physical units

D. It automatically proves customer satisfaction

Correct Answer: B. It may assume constant retention and margin when customer economics change over time

Question 10

Which dashboard practice is most decision-oriented?

A. Displaying every available metric without definitions

B. Changing metric definitions whenever performance falls

C. Linking a small set of defined metrics to thresholds, owners, and actions

D. Reporting point estimates without forecasts or assumptions

Correct Answer: C. Linking a small set of defined metrics to thresholds, owners, and actions