Forecasting, Optimization, and Marketing Decision Models

Introduction

Marketing decisions increasingly require managers to allocate scarce resources under uncertainty. A firm must estimate future demand, determine how much to spend, choose when campaigns should run, assign sales territories, and decide which customer segments deserve investment. These decisions are interdependent: a demand forecast influences the budget; the budget constrains campaign choices; campaign choices affect demand; and the resulting outcomes provide new data for future forecasts.

Marketing analytics addresses this problem through a connected set of methods: time-series forecasting, causal and response modeling, scenario analysis, simulation, optimization, and decision analysis. No method eliminates uncertainty. Rather, each method makes assumptions explicit, quantifies trade-offs, and helps decision makers compare alternatives.

The central managerial principle is that prediction and decision are different tasks. A forecast estimates what may happen under stated conditions. An optimization model recommends what should be done given objectives and constraints. A decision model combines forecasts, preferences, risks, and actions. Confusing these roles can produce technically sophisticated but strategically inappropriate recommendations.

Learning Objectives

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

1. Diagnose marketing time series and select appropriate forecasting methods.
2. Evaluate forecast accuracy using measures appropriate to the business problem.
3. Construct scenarios and simulation models for uncertain demand and marketing response.
4. Formulate response, budget-allocation, scheduling, and territory-design models.
5. Solve and interpret linear, integer, nonlinear, and robust optimization problems.
6. Apply expected-value, utility, sensitivity, and value-of-information reasoning.
7. Identify ethical, governance, data-quality, and implementation risks.

1. The Marketing Decision System

A useful decision system contains five linked elements:

A forecast should specify its horizon, frequency, target variable, conditioning information, and uncertainty. For example, “monthly demand” is incomplete unless the analyst identifies the product, market, channel, unit of measurement, forecast horizon, and whether promotions are assumed to continue.

Marketing decisions commonly involve three forms of uncertainty:

• Aleatory uncertainty: inherent variation, such as random customer purchases.
• Epistemic uncertainty: imperfect knowledge, such as an uncertain advertising-response curve.
• Decision uncertainty: uncertainty created by choosing among actions whose consequences are not fully known.

Separating these forms helps managers determine whether more data, better modeling, or a different risk policy is needed.

2. Time-Series Forecasting for Demand Planning

A time series is an ordered sequence of observations, such as weekly orders or daily website conversions. Its principal components are level, trend, seasonality, cyclical movement, and irregular variation. A conceptual decomposition is:

where Y_t is the observed value at time t, T_t is the underlying level or trend, S_t is a seasonal component, and R_t is the residual. This additive form is suitable when seasonal fluctuations have approximately constant absolute size. A multiplicative form, Y_t = T_t × S_t × R_t, is more suitable when seasonal variation grows with the level. The multiplicative model requires positive values and is often analyzed after logarithmic transformation.

2.1 Data preparation and diagnosis

Before fitting a model, the analyst should:

1. Define the forecasting target and unit.
2. Align calendars, time zones, fiscal periods, and reporting cutoffs.
3. Identify stockouts, returns, cancellations, and delayed transactions.
4. Distinguish missing observations from genuine zero demand.
5. Record promotions, price changes, distribution changes, and competitor events.
6. Partition data into training, validation, and test periods using time order.

Randomly splitting time-series observations can leak future information into the training set. Rolling-origin validation is usually preferable: fit the model on an initial period, forecast the next period, expand the training window, and repeat.

2.2 Baseline and smoothing methods

A baseline is essential. The naïve forecast is:

where Ŷ_t+h|t is the forecast made at time t for h periods ahead. A seasonal naïve forecast uses the observation from the corresponding previous season:

where m is the seasonal period, such as 12 for monthly data with annual seasonality. These baselines are difficult to outperform consistently and provide a meaningful benchmark.

A moving-average forecast with window length k is:

It smooths noise but reacts slowly to changes. Weighted moving averages assign larger weights to more recent observations, with nonnegative weights summing to one.

Simple exponential smoothing is appropriate for a series with level but no systematic trend or seasonality:

where L_t is the estimated level and α controls responsiveness. The one-step forecast is Ŷ_t+1|t = L_t. A high α responds quickly but may overreact to noise; a low value is more stable.

For trend, Holt’s linear method uses level and trend states:

where B_t is the trend estimate and β is the trend-smoothing parameter. Forecasts may become unrealistic over long horizons if a linear trend is extrapolated indefinitely.

For trend and seasonality, Holt–Winters methods update level, trend, and seasonal states. Additive seasonality is suitable for constant seasonal amplitude; multiplicative seasonality is suitable when amplitude scales with the level. Parameters should be estimated or selected using a time-ordered validation procedure rather than chosen solely for in-sample fit.

2.3 Autoregressive and causal models

An autoregressive model relates the current value to prior values. An AR(1) model is:

where c is an intercept, φ is the autoregressive coefficient, and ε_t is an error term with mean approximately zero. A stationary AR(1) process generally requires |φ| < 1. Autoregressive integrated moving-average models can represent autocorrelation, trend after differencing, and moving-average error structure.

Forecasting alone may be inadequate when marketing actions drive demand. A dynamic regression or intervention model can include price, advertising, distribution, promotions, holidays, and competitor actions:

Here, X_t is a vector of other predictors, and Adstock_t represents the accumulated effect of advertising. A simple adstock recursion is:

The coefficient λ describes persistence. This specification is not automatically causal: advertising may be increased precisely when demand is expected to rise. Causal interpretation requires design or identification assumptions, such as randomized experiments, credible natural experiments, instrumental variables, or carefully justified controls.

3. Forecast Evaluation and Uncertainty

Let actual values be Y_t and forecasts be Ŷ_t. Common measures include:

MAE is in the original units and is easily interpreted. RMSE penalizes large errors more strongly. MAPE is scale-free but is undefined when Y_t = 0 and unstable for values near zero. For intermittent demand, scaled measures such as the mean absolute scaled error or measures based on cumulative error may be more suitable.

Accuracy should be evaluated at the level at which decisions are made. A model can be accurate for total monthly demand but poor for individual stores, or accurate on average but systematically biased during promotions. Forecast bias can be monitored with mean error:

Prediction intervals communicate uncertainty. A point forecast without an interval can encourage false precision. Intervals should reflect model error, parameter uncertainty, and—where relevant—uncertainty in future prices, promotions, or media spending.

Forecast combinations often improve robustness. If forecasts from models 1 through J are combined using weights w_j that sum to one, the combination is:

Weights may be equal, validation-based, or estimated under constraints. Combining models is particularly useful when structural uncertainty is substantial.

4. Scenario Analysis and Simulation

A scenario is a coherent set of assumptions about external conditions and managerial actions. Scenarios are not forecasts with arbitrary optimism or pessimism. They should be internally consistent and linked to observable drivers.

A practical scenario table might include:

Scenario analysis is useful when structural changes make historical patterns unreliable. It is less useful when scenarios merely rename numerical outcomes without explaining the assumptions that generate them.

Simulation represents uncertainty by repeatedly sampling uncertain inputs and calculating outcomes. Suppose profit is:

where p is price per unit, c is variable cost per unit, D is demand, F is fixed cost, and M is marketing expenditure. If demand is simulated from a specified distribution, the analyst obtains a distribution of profit rather than one estimate.

For Monte Carlo simulation:

1. Specify probability distributions and dependencies for uncertain inputs.
2. Draw a random vector of inputs.
3. Calculate demand, revenue, cost, and profit.
4. Repeat many times.
5. Summarize expected profit, percentiles, downside probability, and decision comparisons.

Independence should not be assumed casually. Price, demand, media efficiency, and competitor spending may be correlated. A simulation that samples each variable independently can substantially understate risk or create impossible combinations.

5. Response Models and Marketing Contribution

A response model estimates how an outcome changes with controllable inputs. A simple diminishing-returns response curve is:

where x is spend, a is the saturation level, and b > 0 controls the rate of approach to saturation. Its marginal response is:

which decreases as x increases. This diminishing-return property is often plausible, but it must be tested rather than assumed.

A logistic response can represent a lower and upper bound:

where L and U are lower and upper bounds. Such curves may be appropriate for conversion probability or adoption, but parameter interpretation depends on the data and scale.

Incrementality is central. A campaign may generate observed sales that would have occurred without the campaign. The incremental effect is the difference between outcomes under treatment and a credible counterfactual. Randomized holdout tests are often the strongest practical design. Attribution models based only on observed journeys can be useful descriptively but may be biased by selection, targeting, and simultaneous campaigns.

6. Optimization Models for Marketing Allocation

Optimization formalizes a choice among alternatives. A general model is:

Here, x is the decision vector, f(x) is the objective, g_i(x) are constraint functions, b_i are limits, and X defines feasible decisions such as nonnegativity or integrality.

6.1 Linear budget allocation

Suppose channel j receives spend x_j, produces estimated incremental contribution r_j per unit, and has minimum and maximum spend limits. A linear model is:

where B is the budget. This model is transparent but assumes constant returns. If returns diminish, a piecewise-linear approximation or nonlinear model is more realistic.

6.2 Marginal allocation principle

With a smooth objective and a single budget constraint, an interior optimum often equalizes marginal contribution per monetary unit across active channels:

where μ is the shadow value of the budget. This condition applies only under suitable differentiability, feasibility, and interior-solution assumptions. At lower or upper bounds, the marginal condition is replaced by a complementary constraint logic: a channel at its maximum may still have a high marginal return but cannot receive more.

6.3 Campaign scheduling

Scheduling can be modeled with binary variables. Let z_ct = 1 if campaign c runs in period t, and zero otherwise. A simple model may maximize expected contribution:

subject to period capacity, campaign frequency, spacing, and budget constraints. For example, if at most one campaign can run in period t:

If a campaign must run exactly three times:

Binary variables make the model a mixed-integer program. The solution may be computationally harder than a continuous allocation, but it represents operational reality more faithfully.

6.4 Territory design

Territory design commonly balances workload, potential, travel time, service levels, and fairness. Let y_is = 1 indicate assignment of customer i to territory s. Each customer must be assigned once:

Workload constraints can be written as:

where w_i is customer workload and L_s, U_s are lower and upper limits. A territory model may also include geographic contiguity, travel distance, representative skills, and service restrictions. Purely maximizing revenue can create inequitable or impractical territories.

7. Decision Analysis Under Uncertainty

When alternatives have uncertain outcomes, expected monetary value is:

where a is an action, s is a state of nature, P(s) is the probability of that state, and Payoff(a,s) is the resulting payoff. EMV is appropriate when probabilities are credible and the decision maker is risk neutral with respect to the modeled payoff.

Risk-averse managers may use expected utility rather than expected monetary value. With utility function U(Π), select the action maximizing:

A concave utility function represents diminishing utility of additional profit. The certainty equivalent is the guaranteed amount producing the same utility as the risky option. Sensitivity analysis should vary probabilities, response coefficients, costs, and constraints to identify assumptions that determine the recommendation.

The expected value of perfect information is:

It is an upper bound on what perfect information is worth. The expected value of sample information is lower or equal to EVPI and accounts for an actual research design. These concepts help determine whether additional experimentation or market research is economically justified.

8. Integrated Illustrative Case: Northstar Home Devices

Background

Northstar Home Devices sells a subscription-connected air-quality monitor through a website, retail partners, and paid digital media. Monthly demand shows annual seasonality, with stronger sales during winter. The company plans its next six-month media budget and must decide whether to increase digital spend, retailer promotions, or both.

This is an illustrative case, not a report of an actual company.

Problem

The historical data contain a major stockout during one winter month, a price change, and a change in paid-search tracking. A naïve regression of sales on advertising would likely overstate media effectiveness because advertising was increased during the winter peak and because observed sales were constrained by inventory.

Analysis

The analytics team first reconstructs unconstrained demand using inventory availability indicators and flags the stockout. It creates a seasonal naïve benchmark and compares it with a dynamic regression containing price, promotion, holiday indicators, distribution, and an advertising stock variable. Rolling-origin validation is used, and performance is evaluated with MAE and bias at the monthly planning level.

The team then estimates response curves using a randomized regional holdout for digital media. The experiment indicates positive incremental sales but diminishing returns. Retailer promotions are evaluated separately because they affect both volume and margin. The resulting simulation samples demand, media efficiency, and supply availability jointly rather than independently.

Approach and solution

Three scenarios are constructed:

A mixed-integer allocation model selects monthly channel spending and promotion timing. It includes a total budget, minimum retailer support, a maximum digital-spend limit, and a rule preventing two high-discount promotions in consecutive months. The objective is simulated expected contribution margin, not revenue alone. A downside constraint limits the probability that contribution falls below a specified threshold.

Outcome

The initial plan allocated almost all funds to digital media because it used average response rather than incremental response. The revised plan shifted some spending to retailer promotions in months where inventory was strong and digital marginal returns were low. It also reserved budget for the high-uncertainty winter period. The recommendation was not simply the plan with the highest point estimate; it was the plan with the best risk-adjusted contribution under plausible scenarios.

Learning points

1. Stockouts can make observed sales an underestimate of demand.
2. Correlation between advertising and demand does not establish incrementality.
3. The objective should reflect contribution, capacity, and risk—not only sales.
4. Scenario and simulation outputs should be translated into operational rules.
5. A model must be monitored after implementation because response and channel conditions change.

9. Implementation, Governance, and Ethical Issues

A decision model is only as credible as its data and governance. Important controls include versioned data pipelines, documented variable definitions, reproducible code, approval of model changes, and clear ownership of assumptions. Forecasts and optimization results should retain the input data, model version, solver status, and decision rationale.

Fairness matters in targeting and territory design. Optimizing expected profit may systematically deprioritize rural, low-income, older, or protected groups. Analysts should test whether allocation rules create disparate access or exploit vulnerability. Consent, purpose limitation, data minimization, retention controls, and applicable privacy law are relevant when customer-level data are used.

AI-assisted forecasting can improve feature construction, anomaly detection, and model comparison, but it can also amplify leakage, instability, and opacity. Automated recommendations should include human review, monitoring for drift, and safeguards against using sensitive or proxy variables inappropriately. A model that predicts response accurately may still be unacceptable if the action it recommends is coercive, discriminatory, or inconsistent with brand commitments.

10. Best Practices and Limitations

Best practice is to begin with a decision statement: who decides, by when, using which objective, under which constraints? Then establish a transparent benchmark, validate models out of time, quantify uncertainty, and compare model recommendations with managerial knowledge.

Common failure modes include:

• optimizing a biased or non-incremental outcome;
• using future information during training;
• treating a point forecast as a guarantee;
• ignoring capacity, inventory, or operational constraints;
• assuming independent uncertainties;
• maximizing revenue while neglecting margin and customer welfare;
• selecting a complex model that stakeholders cannot audit;
• failing to monitor performance after deployment.

Optimization does not compensate for a poor objective. A mathematically optimal answer to an ethically or economically inappropriate objective is still a poor decision. Similarly, a more accurate forecast is not automatically more valuable if it does not change an action or improve a decision.

Contemporary Developments

Modern marketing decision systems increasingly combine probabilistic forecasts, causal experiments, constrained optimization, and real-time monitoring. Digital platforms permit frequent experimentation, but rapid experimentation increases the risk of multiple testing, short-lived effects, and interference across customers. Privacy-preserving measurement and aggregated lift estimation are becoming more important as individual-level tracking declines.

Digital twins and agent-based simulations can represent interactions among customers, channels, competitors, and capacity. Reinforcement learning may be useful for sequential decisions, such as repeated offers, but it requires careful reward design and safeguards against unsafe exploration. In most managerial settings, transparent hybrid systems—forecasting plus experiments plus optimization—remain more defensible than fully autonomous decision making.

Chapter Summary

Forecasting estimates future demand or response, whereas optimization recommends actions under objectives and constraints. Time-series methods represent level, trend, seasonality, autocorrelation, and external drivers. Baselines, rolling-origin validation, error measures, bias checks, and prediction intervals are essential for responsible evaluation.

Scenario analysis explores coherent alternative environments, while simulation quantifies the distribution of outcomes under uncertain and potentially correlated inputs. Response models describe incremental marketing effects and must be distinguished from descriptive attribution. Optimization models allocate budgets, schedule campaigns, and design territories using continuous, binary, or mixed-integer decisions. Decision analysis adds probabilities, utilities, sensitivity analysis, and the value of information.

Effective marketing analytics integrates statistical validity, operational feasibility, economic objectives, ethical governance, and continuous learning. The strongest recommendation is not necessarily the one with the highest point estimate; it is the one that remains appropriate under uncertainty, can be implemented, and can be monitored and revised.

Key Terms

Forecast: An estimate of a future value conditional on available information and stated assumptions.

Time series: Observations ordered by time.

Seasonality: A recurring pattern associated with a calendar or fixed period.

Exponential smoothing: A forecasting approach that updates states using weighted recent observations.

Adstock: A representation of the carryover or persistence of advertising effects.

Forecast bias: Systematic tendency for forecasts to overpredict or underpredict.

MAE: Mean absolute error, measured in the original units of the target.

RMSE: Root mean squared error, which emphasizes larger errors.

Scenario: A coherent set of assumptions describing a possible environment.

Monte Carlo simulation: Repeated random sampling used to generate an outcome distribution.

Response model: A model relating an action, such as spend, to an outcome.

Incrementality: The change caused by an intervention relative to a credible counterfactual.

Objective function: The quantity an optimization model maximizes or minimizes.

Constraint: A condition limiting feasible decisions.

Mixed-integer program: An optimization model containing continuous and discrete variables.

Shadow price: The marginal value of relaxing a binding constraint.

Expected monetary value: Probability-weighted average payoff.

Expected utility: Probability-weighted utility of uncertain payoffs.

Value of information: The improvement in decision value obtainable from additional information.

Model governance: Processes for documenting, validating, approving, monitoring, and controlling models.

Review Questions

Short Answer Questions

1. Distinguish a forecast from an optimization recommendation.
2. When is a seasonal naïve forecast useful?
3. Why can MAPE be problematic for demand data containing zeros?
4. What does incrementality mean in campaign measurement?
5. State two governance controls for marketing decision models.

Long Answer Questions

1. Compare moving averages, exponential smoothing, and dynamic regression for marketing demand forecasting. Discuss assumptions and limitations.
2. Explain how scenario analysis and Monte Carlo simulation can be combined in a marketing budgeting process.
3. Formulate a constrained campaign-scheduling problem and explain the meaning of its decision variables, objective, and constraints.
4. Discuss why correlation between advertising spend and sales is insufficient to establish causal effectiveness.
5. Explain how risk aversion, sensitivity analysis, and value of information improve decision analysis.

Analytical / Application Questions

1. A retailer has monthly demand with annual seasonality, frequent stockouts, and changing promotion intensity. Design a forecasting and validation process that avoids leakage and distinguishes demand from observed sales.
2. A marketing director must allocate a fixed budget across two channels with diminishing returns, minimum commitments, and uncertain response. Specify an appropriate simulation-optimization workflow and identify the assumptions that require testing.
3. A territory redesign increases expected revenue but creates large workload differences among sales representatives. Propose an optimization objective and constraints that balance revenue, service, travel, and fairness.

Multiple Choice Questions with Answers

Question 1

Which method is most appropriate as a benchmark for monthly demand with stable annual seasonality?

A. Seasonal naïve forecasting

B. Randomly shuffled cross-validation

C. A model using future sales values

D. A forecast based only on average price

Correct Answer: A. Seasonal naïve forecasting

Question 2

In simple exponential smoothing, what does a larger α generally imply?

A. Less responsiveness to recent observations

B. More responsiveness to recent observations

C. Automatic inclusion of annual seasonality

D. A guarantee of unbiased forecasts

Correct Answer: B. More responsiveness to recent observations

Question 3

Why is MAPE problematic when actual demand equals zero?

A. It requires forecasts to be integers

B. It is undefined because the denominator is zero

C. It always produces a negative value

D. It ignores the forecast horizon

Correct Answer: B. It is undefined because the denominator is zero

Question 4

What is the primary purpose of a randomized holdout in campaign measurement?

A. To estimate incremental impact against a counterfactual

B. To maximize the campaign budget

C. To remove all seasonal demand

D. To guarantee that every customer purchases

Correct Answer: A. To estimate incremental impact against a counterfactual

Question 5

In a budget-allocation model, what does a binding budget constraint mean?

A. The budget has no effect on the solution

B. The budget limit is reached at the solution

C. All channels must receive equal spending

D. The objective is necessarily minimized

Correct Answer: B. The budget limit is reached at the solution

Question 6

Which condition is characteristic of an interior optimum with a single smooth budget constraint?

A. All channels receive zero spending

B. Marginal contributions are equal across active channels

C. Every channel receives its maximum limit

D. Forecast errors must be normally distributed

Correct Answer: B. Marginal contributions are equal across active channels

Question 7

What does Monte Carlo simulation primarily provide?

A. A distribution of possible outcomes

B. A proof that the forecast is causal

C. A replacement for all business constraints

D. A single guaranteed outcome

Correct Answer: A. A distribution of possible outcomes

Question 8

Expected monetary value is most directly calculated by:

A. Selecting the largest possible payoff only

B. Averaging payoffs without probabilities

C. Probability-weighting payoffs across states

D. Subtracting forecast error from revenue

Correct Answer: C. Probability-weighting payoffs across states

Question 9

Which statement best describes adstock?

A. A measure of inventory turnover

B. A representation of advertising carryover over time

C. A constraint requiring equal media spend

D. A customer privacy regulation

Correct Answer: B. A representation of advertising carryover over time

Question 10

Why should optimization objectives often use contribution margin rather than revenue alone?

A. Revenue always equals profit

B. Contribution accounts for relevant variable economic costs

C. Margin eliminates all uncertainty

D. Revenue cannot be forecast

Correct Answer: B. Contribution accounts for relevant variable economic costs