Market Sizing, Demand Forecasting and Experimentation

Introduction

Entrepreneurs make decisions before reliable evidence exists. They must estimate how many customers may buy, how frequently they may buy, what price they may accept, and how demand may change over time. These estimates influence product design, financing, hiring, capacity, market entry and investor communication. Yet market size is not a single objective number, and a forecast is not a promise. Both are structured hypotheses about an uncertain future.

Market sizing estimates the potential scale of an opportunity. Demand forecasting estimates the timing and magnitude of likely purchases. Experimentation generates evidence by testing customer behaviour, rather than relying only on stated intentions. Used together, these methods enable disciplined entrepreneurial judgment: sizing frames the opportunity, forecasting supports planning, and experimentation reduces uncertainty.

Learning Objectives

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

1. Distinguish TAM, SAM and SOM and calculate market size using top-down and bottom-up methods.
2. Evaluate the strengths and weaknesses of industry analogues and benchmark data.
3. Apply time-series and judgmental forecasting appropriately.
4. Design concept tests, A/B tests and minimum viable product experiments.
5. Interpret experimental evidence under uncertainty and avoid common analytical errors.
6. Integrate market estimates, forecasts and experiments into entrepreneurial decisions.

1. Market Sizing as an Explicit Set of Assumptions

Market sizing answers a bounded question: how much economic demand could exist for a defined offering, customer group, geography and time period? A credible estimate specifies the unit being counted. “The global wellness market” is less useful than “annual spending by urban Indian firms with 50–500 employees on employee mental-health subscriptions.”

Three related concepts are commonly used:

TAM, SAM and SOM are not automatically three different data points. They are successive filters applied to a clearly defined market. A small SOM does not necessarily imply a poor business; a focused niche can support a profitable venture even when its TAM is modest.

A useful market-sizing statement includes:

• customer segment and inclusion criteria;
• product or service definition;
• geographic and time boundaries;
• unit of demand, such as customers, transactions or subscriptions;
• price or revenue basis;
• assumptions about adoption, replacement and frequency; and
• sources, confidence level and date of evidence.

1.1 Top-Down Sizing

Top-down sizing begins with an external market estimate and narrows it through defensible filters. For example, an analyst may begin with industry revenue, retain the relevant geography, exclude unsuitable customer segments, and estimate the category share corresponding to the proposed solution.

A generic formulation is:

where N is the number of potentially relevant customers, f is annual purchase frequency, and p is revenue per purchase. The units must be consistent: customers × purchases/customer/year × currency/purchase = currency/year.

A filtering formulation can be written as:

where g, s and c are, respectively, the proportions within the target geography, segment and channel or capability boundary. These factors should not be multiplied mechanically when they overlap. For example, geography and segment may already be jointly reported; double counting would understate or distort the result.

Top-down analysis is efficient and useful for strategic context, but it can inherit poor category definitions, optimistic industry reports and inappropriate growth assumptions. It should therefore be treated as a hypothesis requiring bottom-up validation.

1.2 Bottom-Up Sizing

Bottom-up sizing builds from observable operating units. It may estimate the number of target businesses, sales capacity, conversion rate, purchase frequency and price. A subscription example is:

where n is the number of reachable accounts, a is the proportion matching the target profile, r is the conversion or adoption rate, p is monthly price per account, and 12 converts monthly revenue to annual revenue.

Suppose a venture can identify 20,000 reachable small retailers. If 40% fit its technology and payment requirements, 8% adopt within the planning period, and the monthly subscription is ₹1,500, then:

• reachable qualified accounts = 20,000 × 0.40 = 8,000;
• adopting accounts = 8,000 × 0.08 = 640;
• monthly recurring revenue = 640 × ₹1,500 = ₹960,000;
• annual recurring revenue, assuming all accounts remain active for twelve months, = ₹11,520,000.

The retention assumption is important. If customers join gradually or churn, multiplying the final customer count by twelve overstates annual revenue. A cohort schedule or monthly model is more appropriate.

Bottom-up methods are persuasive because assumptions connect to sales territories, customer lists, capacity and observed conversion. Their weakness is possible false precision: a detailed spreadsheet can still be wrong if reachability or conversion rates are speculative.

1.3 Analogues and Triangulation

Analogues use comparable products, companies, regions or historical launches to inform an estimate. A comparable venture may reveal realistic acquisition costs, adoption speed, average contract value or churn. However, “comparable” must be analysed rather than asserted. Relevant dimensions include customer type, price, regulation, distribution, switching costs, product maturity and macroeconomic conditions.

Triangulation combines at least two approaches. For example:

1. Top-down data establish the category boundary.
2. Bottom-up calculations estimate reachable demand.
3. Analogues test whether adoption, price and growth assumptions are credible.
4. Primary research identifies disconfirming evidence.

If the methods produce different results, the discrepancy is diagnostically useful. It may indicate an overly broad category, an unrealistic conversion rate, an omitted competitor or a supply constraint.

2. Demand Forecasting

Market size concerns potential scale; demand forecasting concerns expected demand over a specified horizon. Forecasts should state the horizon, forecast unit, data frequency, assumptions and uncertainty range. A monthly forecast of paid subscriptions is different from a five-year forecast of category revenue.

2.1 Time-Series Forecasting

Time-series methods use historical observations ordered over time. They are most useful when the venture has repeated observations and the underlying environment is reasonably stable.

A simple moving average of order k is:

where D_t is demand observed in period t, F_t+1 is the forecast for the next period, and k is the number of periods averaged. The method gives equal weight to the selected observations and is unsuitable when strong trend or seasonality is ignored.

Exponential smoothing gives more weight to recent data:

where 0 ≤ α ≤ 1 is the smoothing parameter. A larger α responds more rapidly to recent changes; a smaller value produces a smoother forecast. The formula assumes a level series without explicitly modelling trend or seasonality.

For a series with a roughly linear trend, a least-squares model may be used:

where a is the intercept, b is average change per time period, t is time, and ε_t is the unexplained error. Trend extrapolation becomes unreliable when structural changes, competition or capacity limits are likely.

Forecast accuracy should be evaluated on holdout periods rather than only on the data used to fit the model. Common measures include mean absolute error:

where n is the number of evaluated periods. MAE has the same units as demand and is easy to interpret. Percentage measures can be misleading when actual demand is zero or close to zero.

Worked Forecast Example

A venture recorded weekly orders of 100, 120 and 110. A three-period moving-average forecast for week four is:

If actual week-four demand is 130, the absolute error is 20 orders. This single error does not establish whether the model is good; performance must be assessed across multiple periods and compared with alternatives.

2.2 Judgmental Forecasting

Judgmental forecasting is necessary when historical data are absent, the product is novel or a structural break makes past data unreliable. Methods include expert panels, sales-force estimates, customer intention studies, scenario analysis and the Delphi method, in which experts provide repeated, often anonymised estimates with controlled feedback.

Judgment should be structured rather than presented as intuition. A useful process is:

1. define the forecast variable and horizon;
2. collect independent estimates before group discussion;
3. require explicit assumptions and base rates;
4. separate optimistic, base and pessimistic scenarios;
5. record confidence and reasons for disagreement;
6. update the forecast as evidence arrives.

Judgmental methods are vulnerable to anchoring, overconfidence, availability bias, strategic manipulation and social conformity. Combining independent estimates does not remove bias, but it can expose assumptions and reduce dependence on one individual.

2.3 Scenario and Sensitivity Analysis

A point forecast conceals uncertainty. Scenario analysis varies major drivers such as adoption, price, churn, capacity and competitor response. Sensitivity analysis changes one assumption at a time to identify which variables most affect the outcome.

For a subscription business, annual revenue can be represented as:

where C is the average monthly customer base, p is monthly price, and c is the monthly churn rate, provided the expression is being used as a simplified steady-state approximation. It should not be interpreted as an exact cohort model. A complete forecast should model additions and losses by month.

3. Experimentation Under Uncertainty

Sizing and forecasting estimate what might happen; experimentation tests what happens under specified conditions. An experiment is valuable when it changes a decision. Before running one, the entrepreneur should state the decision rule: for example, “Proceed to a paid pilot if at least 20% of qualified prospects complete payment and the estimated gross margin remains positive.”

Experiments should distinguish:

• desirability: do customers want the solution?
• viability: will revenue exceed relevant costs?
• feasibility: can the team deliver reliably?
• usability: can customers understand and use it?
• scalability: does performance persist at larger volume?

3.1 Concept Tests

A concept test presents a proposed solution before full development. It may use a written proposition, storyboard, mock-up, landing page or prototype. Strong concept tests measure more than verbal enthusiasm. They examine comprehension, relevance, willingness to try, willingness to pay, objections and the next action taken.

A concept-test instrument should:

• recruit respondents who resemble the intended customer;
• present a standardised concept;
• distinguish unaided reactions from prompted ratings;
• ask about current alternatives and behaviour;
• include a behavioural commitment, such as joining a waitlist or scheduling a demo;
• record reasons for rejection as well as acceptance.

Stated purchase intention is not equivalent to purchase behaviour. Social desirability, hypothetical bias and courtesy bias can produce inflated results. Asking about a recent comparable purchase and testing a real commitment improves evidence quality.

3.2 A/B Testing

An A/B test randomly assigns eligible users to version A or version B, changes one primary treatment where possible, and compares a predefined outcome. Randomisation helps balance observed and unobserved characteristics in expectation. It does not guarantee balance in a small sample and cannot correct for biased eligibility or interference between users.

For a binary conversion outcome, the observed conversion rate is:

where x is the number of conversions and n is the number of eligible observations. The difference between treatment and control rates is:

A test should define its primary metric, minimum practically important effect, duration, stopping rule and guardrail metrics before launch. Repeatedly checking results and stopping when a preferred result appears increases false-positive risk. Multiple outcomes and many variants also require adjustment or cautious interpretation.

Statistical significance is not the same as managerial importance. A tiny improvement may be statistically detectable but economically irrelevant; a meaningful improvement may be uncertain because the sample is small. Results should therefore be reported with an estimated effect, uncertainty interval where appropriate, practical threshold and business impact.

3.3 Minimum Viable Product Experiments

An MVP is not simply the smallest product. It is the smallest credible system that tests a critical business or customer assumption. It may be a concierge service, manual back-office process, clickable prototype, paid pilot or single-feature product.

A good MVP identifies:

1. the riskiest assumption;
2. the customer and context in which it matters;
3. the observable behaviour that would support or weaken it;
4. the smallest ethical test;
5. a threshold for persevere, pivot or stop;
6. the learning that will transfer to the next iteration.

MVPs can create misleading evidence. Early users may be unusually enthusiastic, founder involvement may inflate service quality, and manual delivery may not scale. The experiment should document which elements are temporary and which are intended to represent the eventual business.

4. Integrating Estimates and Evidence

A practical evidence architecture links assumptions to methods. A founder might maintain an assumption register with the following fields:

The evidence hierarchy is not absolute, but observed behaviour under realistic conditions is generally stronger than unprompted opinion, which is stronger than a founder’s assumption. A paid renewal is usually more informative than a click, although it may involve a small and unrepresentative sample.

Bayesian reasoning provides a useful conceptual discipline. Begin with a prior belief, observe evidence, and update the belief according to the reliability and diagnostic value of that evidence. Entrepreneurs need not calculate formal posterior probabilities in every case; they should nevertheless ask: How likely was this evidence if the assumption were true? How likely would it be if the assumption were false? What alternative explanations exist?

5. Illustrative Case: FreshRoute Institutional Produce Platform

Background

FreshRoute is an illustrative venture proposing a platform that connects small institutional kitchens with local produce suppliers. The founders initially claimed a large national market based on total food-service spending.

Problem

The initial estimate was top-down and mixed all restaurants, hotels and institutional kitchens. It did not distinguish fresh produce from total food spending, nor account for the venture’s limited delivery geography. The forecast also assumed immediate adoption by every qualified kitchen.

Analysis

The team created a bottom-up model for three launch districts. It identified 2,400 institutional kitchens, of which 900 had sufficient order frequency and digital purchasing processes. Interviews suggested that 30% had a recurring procurement problem severe enough to consider a new supplier. The proposed monthly platform fee was ₹2,000.

Potential early adopters were therefore estimated as:

If 15% converted during the first year, expected customers would be 40.5, rounded to approximately 40 for planning. Subscription revenue would be:

This excluded transaction commissions, delivery revenue and churn; it was deliberately a conservative subscription-only estimate.

The founders then ran a concept test with procurement managers and a landing page that offered a paid sourcing pilot. Interviews produced positive reactions, but only 9 of 60 qualified visitors requested a pilot. Six paid an initial deposit. The team interpreted this as evidence of interest, not proof of scalable demand.

For the MVP, FreshRoute manually matched ten kitchens with suppliers for eight weeks. The primary success criteria were repeat ordering by at least six kitchens, fulfilment above 95% and contribution margin above zero after delivery costs. Seven kitchens placed repeat orders, fulfilment was 96%, but contribution margin was negative because routes were inefficient.

Approach and Outcome

The venture did not immediately scale customer acquisition. It narrowed the launch geography, introduced minimum order values, grouped deliveries by route and tested a higher-value procurement dashboard. The market estimate was reduced, but operational learning improved the forecast. The founders concluded that demand existed, while the original delivery model was not economically viable.

Learning Points

• A broad industry number can conceal a much smaller reachable market.
• Paid and repeated behaviour is stronger evidence than favourable interviews.
• Demand and unit economics must be tested together.
• An experiment can validate customer need while falsifying the delivery model.
• Negative results are useful when the decision rule and learning objective were explicit.

6. Challenges, Ethics and Governance

Market estimates can mislead investors, employees and customers when assumptions are hidden or numbers are presented as facts. Professional practice requires transparent sources, dated data, clear units, sensitivity ranges and separation of observed results from projections.

Customer experiments must respect informed consent, privacy and proportionality. Data should be collected for a legitimate purpose, minimised, secured and retained only as long as necessary. Personalisation tests may create discriminatory outcomes if algorithms target vulnerable groups or allocate prices unfairly. Experiments involving essential services, financial decisions, health claims or children require heightened review and compliance with applicable laws.

Entrepreneurs should not manipulate control groups, conceal material risks, fabricate scarcity or use dark patterns to obtain consent. A statistically successful test is not ethically acceptable if it relies on deception or causes unreasonable harm. Governance should include access controls, experiment logs, version histories, pre-specified metrics and a process for stopping harmful tests.

7. Technology and Contemporary Practice

Digital analytics, cloud experimentation platforms and machine-learning forecasting can accelerate learning. Automated models are useful when data are sufficient, stable and representative. They are not substitutes for market definition or causal reasoning. Machine-learning forecasts may overfit historical patterns, reproduce biased data and fail during structural change.

Modern ventures increasingly use sequential experimentation, feature flags, synthetic data, cohort analysis and causal-inference methods. These tools require methodological discipline. A dashboard showing correlation does not establish that a product change caused the outcome. Privacy-preserving measurement, responsible data governance and human review are increasingly important as experimentation becomes automated.

Best-Practice Checklist

Before presenting a market estimate or forecast, ask:

1. What exactly is being counted, and in what units?
2. Are TAM, SAM and SOM defined by observable boundaries?
3. Which assumptions drive the result most strongly?
4. Has a top-down estimate been triangulated bottom-up?
5. Does the forecast include trend, seasonality, capacity and competition where relevant?
6. Are uncertainty ranges and scenarios shown?
7. What customer behaviour would change the decision?
8. Is the experiment randomised or otherwise designed to reduce alternative explanations?
9. Are metrics, stopping rules and ethical safeguards specified in advance?
10. Has the evidence been updated after real customer behaviour?

Chapter Summary

Market sizing, forecasting and experimentation are complementary methods for entrepreneurial decision-making under uncertainty. TAM, SAM and SOM must be defined by customer, product, geography, time and economic unit. Top-down estimates provide category context, while bottom-up estimates connect assumptions to reachable customers, price, frequency and capacity. Analogues and triangulation improve credibility but do not eliminate uncertainty.

Time-series methods such as moving averages and exponential smoothing are useful when historical patterns are informative; judgmental and scenario methods are necessary when products or markets are novel. Forecasts should be evaluated on holdout data and accompanied by sensitivity analysis.

Concept tests explore desirability, A/B tests compare controlled alternatives, and MVP experiments test critical assumptions with minimal investment. The strongest evidence usually involves realistic, repeated or paid behaviour, but every experiment must consider sample bias, measurement error, practical significance and ethics. Professional entrepreneurs make assumptions explicit, define decision thresholds, record uncertainty and update plans as evidence accumulates.

Key Terms

TAM: Total addressable market under an unrestricted but defined market boundary.

SAM: Serviceable available market reachable with the current offering and constraints.

SOM: Serviceable obtainable market realistically attainable within a specified period.

Top-down sizing: Estimating a target market by filtering an external aggregate estimate.

Bottom-up sizing: Estimating opportunity from operating units such as customers, price and frequency.

Analogue: A comparable market, product or company used to inform assumptions.

Demand forecast: An estimate of expected demand over a stated time horizon.

Time series: Observations recorded in chronological order.

Moving average: A forecast formed from the arithmetic mean of recent observations.

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

Scenario analysis: Evaluation of outcomes under alternative coherent assumptions.

Sensitivity analysis: Assessment of how changing an input affects an outcome.

Concept test: Research evaluating a proposed offering before full development.

A/B test: A randomised comparison of two versions against a defined outcome.

Conversion rate: The proportion of eligible observations completing a specified action.

MVP: The smallest credible experiment or product system that tests a critical assumption.

Guardrail metric: A measure used to detect harmful or unacceptable side effects.

Statistical significance: Evidence that an observed result is unlikely under a specified null model, subject to assumptions.

Practical significance: The managerial or economic importance of an effect.

Triangulation: Combining independent methods or sources to assess the same question.

Review Questions

Short Answer Questions

1. Distinguish TAM, SAM and SOM.
2. Why can bottom-up sizing be more operationally useful than top-down sizing?
3. State the main assumption behind a simple moving-average forecast.
4. Why is stated purchase intention weaker evidence than a paid pilot?
5. What is the purpose of a guardrail metric in an A/B test?

Long Answer Questions

1. Explain how a start-up should triangulate top-down, bottom-up and analogue-based market estimates.
2. Compare time-series and judgmental forecasting, including their appropriate applications and limitations.
3. Describe the design of a rigorous A/B test from hypothesis through interpretation.
4. Explain how an MVP can test desirability, feasibility and viability without becoming a full-scale product.
5. Discuss ethical, privacy and governance requirements for entrepreneurial experimentation.

Analytical / Application Questions

1. A venture identifies 15,000 potential accounts. Forty percent fit its segment, 10% are expected to adopt, and the monthly price is ₹2,500. Calculate the simplified first-year recurring revenue and identify two assumptions that could make the estimate unreliable.
2. Weekly demand for three periods is 80, 100 and 90 units. Calculate the three-period moving-average forecast for the next period. Explain when this forecast would be unsuitable.
3. Design an MVP experiment for a digital procurement service. Specify the riskiest assumption, primary metric, success threshold, ethical safeguard and persevere-or-pivot decision rule.

Multiple Choice Questions with Answers

Question 1

Which statement best defines SOM?

A. The revenue of every firm in a global category

B. The portion of the market reachable with the current product and geography

C. The realistically obtainable market within a specified period

D. The number of customers who have heard of the product

Correct Answer: C. The realistically obtainable market within a specified period

Question 2

A venture has 10,000 reachable accounts, of which 20% fit its target segment, and expects 5% of those to adopt at a monthly price of ₹1,000. What is the simplified annual recurring revenue?

A. ₹100,000

B. ₹1,000,000

C. ₹1,200,000

D. ₹12,000,000

Correct Answer: C. ₹1,200,000

Question 3

What is a major weakness of a top-down market estimate?

A. It cannot use external data

B. It may inherit broad or inappropriate category definitions

C. It always produces a smaller estimate than bottom-up sizing

D. It requires historical customer-level conversion data

Correct Answer: B. It may inherit broad or inappropriate category definitions

Question 4

In exponential smoothing, increasing α generally causes the forecast to:

A. Respond more strongly to recent observations

B. Ignore recent observations

C. Become identical to a moving average of all history

D. Remove all forecast error

Correct Answer: A. Respond more strongly to recent observations

Question 5

A three-period moving-average forecast uses demands of 80, 100 and 90 units. What is the forecast for the next period?

A. 85 units

B. 90 units

C. 95 units

D. 270 units

Correct Answer: B. 90 units

Question 6

What is the principal purpose of randomisation in an A/B test?

A. To guarantee a large sample

B. To balance potential confounding factors in expectation

C. To ensure the treatment always wins

D. To eliminate the need for a predefined metric

Correct Answer: B. To balance potential confounding factors in expectation

Question 7

Which evidence most directly demonstrates willingness to pay?

A. A respondent says the idea is interesting

B. A respondent gives a high concept rating

C. A qualified customer makes a paid pilot deposit

D. A social-media user likes an advertisement

Correct Answer: C. A qualified customer makes a paid pilot deposit

Question 8

Which statement about statistical significance is correct?

A. It proves that an effect is economically important

B. It guarantees that the result will replicate exactly

C. It indicates evidence against a specified null model under stated assumptions

D. It removes the need to examine practical impact

Correct Answer: C. It indicates evidence against a specified null model under stated assumptions

Question 9

What is a guardrail metric?

A. A metric used only to calculate TAM

B. A measure monitoring harmful or unacceptable side effects

C. A substitute for randomisation

D. A forecast of total industry revenue

Correct Answer: B. A measure monitoring harmful or unacceptable side effects

Question 10

Why should a market estimate include sensitivity analysis?

A. To conceal uncertain assumptions

B. To identify how changes in key inputs affect the result

C. To guarantee the highest possible forecast

D. To replace all primary research

Correct Answer: B. To identify how changes in key inputs affect the result