Forecasting, Budget Allocation and Optimization

Introduction

Campaign measurement becomes strategically valuable when it moves beyond reporting what happened and supports decisions about what should happen next. Forecasting estimates likely outcomes under alternative spending plans; budget allocation determines how resources are distributed across channels, audiences, markets and time; optimization selects the feasible plan that best meets an objective.

In advertising, these activities are difficult because response is rarely linear. The first monetary units invested in a channel may reach highly responsive consumers, whereas later units may encounter frequency saturation, weaker audiences, higher auction prices or limited inventory. Consequently, a channel can continue to produce conversions while its incremental return becomes unattractive. The central managerial question is not “Which channel has the highest average ROI?” but “Where will the next unit of budget create the greatest incremental value, subject to business and operational constraints?”

This chapter develops response curves, marginal ROI, forecasting methods, simulation and constrained optimization. It also considers pacing, uncertainty, attribution limitations, privacy, governance and the practical judgment required to turn analytical outputs into responsible media decisions.

Learning Objectives

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

1. Explain why advertising response is nonlinear and represent it using response curves.
2. Distinguish average ROI from marginal ROI and use marginal analysis for allocation.
3. Construct basic forecasts using historical performance, response models and scenario assumptions.
4. Apply simulation to represent uncertainty in campaign outcomes.
5. Formulate budget allocation as a constrained optimization problem.
6. Design pacing rules and evaluate alternative campaign scenarios.
7. Critically assess model assumptions, attribution limitations, ethical risks and governance requirements.

1. The Economics of Campaign Response

A campaign response function maps investment to an outcome such as incremental sales, conversions, qualified leads, reach or profit. Let S_i denote spend in channel i, and let Y_i(S_i) denote the expected incremental outcome generated by that spend. A simple linear model is:

Here, a_i is baseline response independent of the measured spend, and b_i is the constant response per monetary unit. This model is easy to interpret but often unrealistic: it implies that every additional dollar produces the same result indefinitely.

A more realistic diminishing-returns model is the exponential saturation function:

V_i is the maximum attainable response under the model, k_i is the rate at which response approaches that maximum, and e is the base of natural logarithms. Spend and response must be expressed in compatible units, such as dollars and incremental conversions. The model assumes a stable environment, a nonnegative spend level, a fixed maximum response and a response that rises monotonically but at a decreasing rate.

The first derivative gives the marginal response:

This derivative is positive but declines as spend increases. The second derivative is:

The negative second derivative confirms concavity, or diminishing marginal response, under the model’s assumptions.

Other useful response functions include:

In the Hill function, V is the asymptotic maximum, K is the spend at half-maximum response, and h controls curve shape. When h>1, the function can represent a threshold or “learning” region before saturation; when h<1, it reflects immediate diminishing returns. The function is appropriate only when sufficient data support its extra flexibility.

A response curve should not be confused with an attribution report. An attribution report assigns observed outcomes to touchpoints according to a rule. A response curve attempts to estimate how outcomes change when investment changes, ideally using experimental or quasi-experimental evidence. The latter is a causal question and is generally more demanding.

2. ROI, Incremental ROI and Marginal ROI

Return on investment is commonly written as:

where G is return, contribution or value generated, and S is spend. Both must use the same monetary basis. If G is revenue while S is advertising cost, the result is revenue ROI, not profit ROI. If contribution margin is used for G, the result has a more appropriate economic interpretation.

Average ROI at spend S for a channel is:

where M(S) is incremental monetary value. Marginal ROI for a small additional spend change is:

If the response is measured in conversions and each conversion has contribution value v, then M(S)=vY(S) and:

This is a value-per-dollar measure. A marginal ROI greater than 1 means that an additional dollar is expected to generate more than one dollar of contribution value, assuming the same measurement basis and no omitted costs.

Average and marginal ROI can lead to different decisions. A mature channel may have a high average ROI because its initial spend was efficient, while its next dollar has a low marginal ROI. Conversely, a developing channel may have a modest average ROI but a high marginal ROI because it has not yet reached scale.

Worked Example: Average versus Marginal ROI

Suppose Channel A has generated contribution value of $180,000 from $100,000 of spend. Its average ROI is:

A response model estimates that the next $10,000 would generate $11,000 of contribution value. Its incremental or marginal ROI is:

The channel remains profitable at the margin, but its historical 80% ROI should not be used to forecast the next $10,000. Allocation should compare the 10% marginal return with the marginal return available elsewhere.

3. Forecasting Campaign Outcomes

A forecast is a conditional statement about expected future outcomes. It is not a promise. A credible forecast specifies the time horizon, population, outcome, data source, assumptions, uncertainty and intervention being considered.

3.1 Forecasting workflow

A professional forecasting process generally includes:

1. Define the decision. For example, determine next quarter’s paid-media allocation while maintaining a minimum level of reach.
2. Define the outcome. Possible outcomes include incremental contribution, new customers, qualified leads or profit.
3. Assemble and audit data. Check spend, impressions, reach, frequency, conversions, prices, seasonality, promotions and tracking changes.
4. Separate baseline from media-generated response. Baseline demand may arise from brand equity, distribution, pricing, sales activity or macroeconomic conditions.
5. Estimate response relationships. Use experiments, geo-tests, panel data, time-series models or carefully controlled observational methods.
6. Generate a base case and alternatives. State assumptions for prices, inventory, audience size, creative quality, competitors and timing.
7. Quantify uncertainty. Produce intervals or distributions rather than a single falsely precise number.
8. Validate and monitor. Compare forecast with actual results and update the model.

3.2 Basic forecast decomposition

A simple planning identity is:

For multiple channels:

where B is baseline outcome and Y_i(S_i) is the incremental response attributed causally to channel i. This additive form assumes that channel effects do not interact. If channels reinforce or cannibalize one another, interaction terms may be required:

Here, γᵢⱼ represents an interaction effect. A positive value suggests synergy; a negative value suggests overlap or cannibalization. Interaction models require substantial data and should not be added merely because they appear theoretically attractive.

3.3 Forecast error and calibration

Let A_t be actual outcome and F_t be forecast outcome for period t. Forecast error is:

Mean absolute error is:

MAE is expressed in the outcome’s units and is easy to interpret. Mean absolute percentage error is:

MAPE is unsuitable when actual values are zero or close to zero and can overemphasize errors in small denominators. Forecast evaluation should therefore consider scale, business impact and systematic bias, not one metric alone.

A forecast can be statistically accurate but strategically misleading if it predicts attributed conversions rather than incremental conversions. Calibration should test both numerical accuracy and decision usefulness.

4. Budget Allocation as Marginal Optimization

Suppose a firm allocates total budget B across n channels. The basic problem is:

The objective is to maximise total incremental monetary value. This formulation assumes separable channel responses, a fixed budget, and no other constraints.

For an interior optimum with smooth response functions, the Lagrangian is:

The first-order condition is:

Thus, at the optimum, all active channels have equal marginal value, represented by λ. If one channel has a higher marginal value than another, shifting a small amount of budget toward the higher-value channel improves the objective, provided constraints permit the shift. This equal-marginal principle is valid for differentiable concave response functions and a feasible interior solution. Boundaries, discrete buying units and non-concave responses require additional treatment.

Worked Example: Two-Channel Allocation

Assume contribution-value response functions are:

Spend is measured in thousands of dollars and value in thousands of dollars. The available budget is 100, so S_A+S_B=100. Marginal values are:

At an interior optimum:

Rearranging gives √S_A/√S_B = 4/3, so S_A/S_B=16/9. Combining this with S_A+S_B=100 gives:

The allocation is therefore $64,000 to A and $36,000 to B. The common marginal value is 60/8=7.5 thousand dollars of contribution value per additional thousand dollars of spend for A, and 45/6=7.5 for B. This equality is the economic rationale for the allocation.

In practice, media buying is not infinitely divisible. Minimum commitments, insertion orders, audience packages and auction dynamics create discrete or uncertain spending units. A planner may therefore use a grid search, linear programming approximation, mixed-integer optimization or a simulation-based search rather than relying only on calculus.

5. Constraints, Objectives and Practical Optimization

A useful optimization model must represent actual managerial requirements. Common constraints include:

• total budget and channel-specific minimums or maximums;
• minimum reach or effective frequency;
• brand-safety exclusions;
• geographic, audience or legal restrictions;
• inventory capacity;
• working-media and production budgets;
• minimum presence in strategic channels;
• cash-flow or weekly pacing limits;
• sustainability or supplier-diversity requirements.

A general constrained formulation is:

Here, U is the chosen utility function, l_i and u_i are lower and upper spend bounds, and g_k represents a requirement such as reach with threshold r_k. Utility may be contribution profit, expected customer lifetime value, a weighted combination of sales and reach, or a risk-adjusted objective.

The objective must reflect the decision horizon. Maximising immediate conversions may conflict with long-term brand value. A multi-objective function could be:

where P is incremental profit, R is a reach or brand metric, L is long-term value, and α,β,γ are explicit managerial weights. The weights should be documented because changing them can materially change the recommended media mix.

Shadow prices and managerial interpretation

The multiplier λ associated with the budget constraint is a shadow price. Under suitable regularity conditions, it approximates the improvement in the objective resulting from one additional unit of budget. A high shadow price indicates that budget is restrictive; a low value suggests that additional budget would produce limited benefit under the model. Shadow prices are decision aids, not guaranteed financial returns, especially when model uncertainty is high.

6. Simulation and Scenario Analysis

Point forecasts conceal uncertainty. Simulation represents uncertain inputs by probability distributions and repeatedly calculates outcomes. For example, conversion rate, cost per thousand impressions, response-curve parameters and incremental lift may all vary.

Let simulated outcome in run r be:

where θ_i^(r) is a draw of uncertain channel parameters in simulation run r. After R runs, the expected outcome is:

A simulation-based 90% interval can be estimated using the 5th and 95th percentiles of the simulated outcomes. This is a predictive interval, not automatically a confidence interval for a fixed parameter. Its interpretation depends on whether the simulation includes parameter uncertainty, future process variation, or both.

A practical Monte Carlo process is:

1. Specify distributions from experiments, historical residuals or expert judgment.
2. Preserve meaningful correlations; for example, auction cost and conversion rate may move together.
3. Draw parameter values and calculate each channel’s outcome.
4. Apply constraints and record profit, conversions, reach and risk measures.
5. Repeat many times.
6. Compare scenarios using expected value, downside percentiles and probability of meeting objectives.

Scenario comparison

A scenario table can communicate uncertainty more effectively than a single forecast:

Decision-makers should examine the distribution of outcomes. A plan with the highest expected value may be rejected if its probability of losing money is unacceptable. Risk-adjusted utility can be represented as:

where E[P] is expected profit, Var(P) is profit variance and ρ≥0 expresses risk aversion. This expression is a simplified mean–variance objective and is most defensible when variance is a meaningful risk measure and the objective is approximately quadratic. It should not be treated as universally optimal.

7. Pacing and Dynamic Reallocation

An annual or quarterly optimum may fail operationally if spend is delivered too quickly or too slowly. Pacing translates a strategic allocation into time-based controls. If planned spend is S_total over T periods, a simple even target is:

Even pacing is not always optimal. Demand seasonality, promotional windows and auction prices may justify front-loading or back-loading. A dynamic pacing rule can compare actual cumulative spend with target cumulative spend and adjust bids, budgets or inventory commitments.

A useful control metric is pacing ratio:

A ratio above 1 indicates overspending relative to the plan; below 1 indicates underspending. Pacing ratio alone does not establish performance. A campaign may underspend because quality controls are working, or overspend because high-value inventory is temporarily available.

Dynamic reallocation should use updated marginal ROI, not merely delivery gaps. A channel that is behind pace but has poor incremental economics should not automatically receive more money. Reallocation should also consider learning periods, creative fatigue, attribution latency and minimum viable spend.

8. Applied Case: Northstar Home Appliances

The following is an illustrative case designed for learning purposes; it is not a report of an actual company.

Background

Northstar, a regional home-appliance retailer, has a $600,000 six-week media budget. Its historical dashboard shows the following average revenue ROIs: search, 4.2; paid social, 2.8; online video, 1.9; and connected television, 1.6. Management initially proposes allocating most funds to search because it has the highest reported ROI.

Problem

The search team reports that impression share is already high for core product terms and that additional bids are moving into more expensive, less qualified auctions. Social frequency is also rising. The company wants to maximise contribution profit while retaining sufficient upper-funnel reach.

Analysis

The analytics team conducts a geo-experiment and estimates diminishing response curves. It converts outcomes to contribution value rather than revenue and models uncertainty in prices and lift. The team finds:

• Search has the highest historical average ROI but a relatively low marginal ROI at its current spend.
• Paid social has moderate current efficiency and retains useful marginal capacity among prospecting audiences.
• Video has lower short-term attributed ROI but produces measurable incremental branded-search and store visits.
• Connected television has wide uncertainty and is valuable only if minimum reach is required.

The team formulates the objective as expected contribution profit, with a minimum reach constraint and channel bounds reflecting operational commitments. It simulates 10,000 combinations of response parameters and media prices.

Approach and solution

Three plans are compared:

Plan B shifts budget away from mature search auctions, maintains a controlled search presence, increases paid social within its efficient range, and funds video to satisfy the reach constraint. Connected television receives a limited test budget rather than the largest allocation. The plan also reserves 8% of the total budget for mid-flight reallocation.

Outcome

After three weeks, search costs rise above the base-case assumption while social response remains close to forecast. The team reallocates part of the reserve to social and high-intent retail media, but does not eliminate video because the reach requirement remains binding. The final evaluation uses geo-level incremental sales and contribution margin rather than platform-reported conversions alone.

Learning points

1. Historical average ROI is not a reliable guide to the next dollar.
2. A lower-attributed-ROI channel can be strategically valuable when it contributes incremental reach or assists demand.
3. Constraints encode business strategy; removing them changes the question being answered.
4. Reserves and pacing rules allow a plan to respond to new evidence.
5. Simulation makes downside risk visible and supports more robust decisions than a single deterministic forecast.

9. Technology, AI and Implementation Architecture

Modern planning systems combine clean data pipelines, experimentation platforms, media APIs, forecasting models and optimization engines. AI can assist with parameter estimation, anomaly detection, creative classification and scenario generation. However, an AI-generated recommendation is not automatically causal, unbiased or economically valid.

A robust implementation separates four layers:

1. Measurement layer: collects spend, exposure, outcomes, experiments and business controls.
2. Modelling layer: estimates baselines, response curves, uncertainty and interactions.
3. Decision layer: applies objectives, constraints, optimization and scenario analysis.
4. Governance layer: records data provenance, approvals, model versions, overrides and post-campaign validation.

Models should be version-controlled and accompanied by assumption registers. Human review is essential when recommendations affect sensitive audiences, regulated products, employment-related advertising or substantial expenditure. Automated bidding should have spending caps, anomaly alerts, stop-loss rules and escalation procedures.

10. Limitations, Ethics and Governance

Identification and attribution

Observed association is not incremental effect. Users exposed to advertising may already be more likely to purchase. Holdout tests, geo-experiments, synthetic controls and econometric designs can improve identification, but each has assumptions. A model should state whether its estimates are descriptive, predictive or causal.

Data quality and privacy

Budget models may combine customer, device, location and behavioural data. Organisations should apply purpose limitation, data minimisation, retention controls, access restrictions and applicable privacy law. Personally identifiable information should not be included merely because it is available. Consent and legitimate-use requirements vary by jurisdiction and campaign context.

Fairness and exclusion

Optimization can favour audiences that are cheaper to reach or more likely to convert, potentially excluding protected groups from opportunities or exposing vulnerable groups to harmful persuasion. Teams should assess delivery and outcome disparities, document targeting logic and prohibit sensitive inferences where inappropriate.

Model risk

A response curve extrapolated beyond observed spend can produce implausible recommendations. Structural market changes, privacy-related signal loss, platform policy changes, competitor actions and creative fatigue can invalidate historical relationships. Sensitivity analysis, out-of-sample validation and periodic recalibration are therefore necessary.

Transparency and accountability

Managers should be able to explain why a budget was allocated, which assumptions drove the result and what evidence would justify changing course. A sophisticated optimizer without auditability can weaken, rather than improve, governance.

11. Best-Practice Checklist

• Define the business value metric before optimising.
• Use incremental outcomes where the decision is causal.
• Estimate and compare marginal, not only average, returns.
• Avoid extrapolating response curves outside observed evidence without explicit scenarios.
• Include channel, time, inventory and reach constraints.
• Preserve correlations in simulation inputs.
• Report expected outcomes and downside risk.
• Reserve budget for learning and controlled reallocation.
• Validate forecasts with both statistical and business measures.
• Document model versions, assumptions, overrides and approvals.
• Monitor pacing, saturation, frequency, quality and marginal ROI during delivery.
• Apply privacy, fairness, brand-safety and regulatory controls.

Chapter Summary

Forecasting and optimization convert campaign analytics into forward-looking resource decisions. Response curves represent how outcomes change with spend and usually incorporate diminishing returns. Average ROI describes historical or cumulative efficiency, whereas marginal ROI evaluates the next unit of investment and is the appropriate basis for reallocating budget when response is nonlinear.

A basic allocation problem maximises total incremental value subject to a budget constraint. For smooth concave functions, the optimal interior solution equalises marginal value across active channels. Real campaigns require additional constraints involving reach, inventory, minimum presence, timing, legal requirements and strategic objectives. Forecasts should distinguish baseline from incremental outcomes, state assumptions and quantify uncertainty.

Simulation allows planners to compare distributions of outcomes rather than relying on a single estimate. Pacing and reserve budgets support dynamic control when prices, response or inventory change. Finally, causal identification, privacy, fairness, transparency and model governance determine whether an analytically attractive recommendation is responsible and actionable.

Key Terms

Response curve: A function describing expected outcome as investment changes.

Diminishing returns: A condition in which each additional unit of spend produces less incremental outcome.

Marginal response: The change in expected outcome from a small increase in spend.

Marginal ROI: Incremental value or return generated by the next unit of investment.

Average ROI: Cumulative return relative to cumulative spend.

Incrementality: The outcome caused by advertising beyond what would otherwise have occurred.

Baseline: Expected outcome in the absence of the campaign or treatment being evaluated.

Forecast: A conditional estimate of a future outcome.

Scenario: A coherent set of assumptions used to examine a possible future.

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

Constraint: A condition limiting feasible decisions.

Objective function: A mathematical representation of what an optimization seeks to maximise or minimise.

Shadow price: The estimated value of relaxing a binding constraint by one unit.

Pacing: Managing delivery of budget and activity over time.

Saturation: A state in which additional investment produces very small incremental response.

Interaction effect: A situation in which the effect of one channel depends on another channel’s investment.

Predictive interval: A range describing plausible future outcomes under stated assumptions.

Model governance: Processes controlling model documentation, validation, use and accountability.

Review Questions

Short Answer Questions

1. Why can average ROI be misleading when allocating the next unit of campaign budget?
2. What does a negative second derivative indicate about a response curve?
3. Distinguish a forecast from a causal estimate of incrementality.
4. What is the purpose of a pacing ratio?
5. Name two constraints that may be included in a media optimization problem.

Long Answer Questions

1. Explain the economic logic of equalising marginal ROI across active channels. State the assumptions under which this principle is valid.
2. Compare linear, exponential saturation and Hill response functions. Discuss when each may be useful and what risks arise from over-flexible modelling.
3. Describe a complete campaign forecasting workflow, from decision definition through post-campaign validation.
4. Explain how Monte Carlo simulation can improve budget decisions and discuss the importance of correlations among uncertain inputs.
5. Discuss the ethical and governance issues involved in using automated optimization and AI for advertising allocation.

Analytical / Application Questions

1. A channel has generated $240,000 of contribution value from $150,000 of spend. The next $20,000 is expected to generate $18,000 of contribution value. Calculate its average ROI and marginal ROI, and explain which figure should guide an incremental allocation decision.
2. A planner has a fixed budget across three channels whose marginal values are currently 8, 5 and 3 contribution dollars per dollar spent. Explain the direction of a small reallocation and identify what information is needed before making a final allocation.
3. Design a simulation-based evaluation of two competing media plans. Specify at least four uncertain inputs, one performance objective, two risk measures and one governance control.

Multiple Choice Questions with Answers

Question 1

Which measure is most appropriate for deciding where the next dollar of budget should be invested when response is nonlinear?

A. Historical average ROI

B. Marginal ROI

C. Total impressions

D. Cumulative reach

Correct Answer: B. Marginal ROI

Question 2

In the exponential response model Y=V(1−e^(−kS)), what does V represent?

A. The maximum response approached by the model

B. The cost per impression

C. The marginal response at every spend level

D. The campaign duration

Correct Answer: A. The maximum response approached by the model

Question 3

For a concave, differentiable allocation problem with only a binding total-budget constraint, an interior optimum generally requires:

A. Equal average ROI across all channels

B. Equal marginal value across active channels

C. Equal impressions across all channels

D. Equal reach across all channels

Correct Answer: B. Equal marginal value across active channels

Question 4

A campaign generates $12,000 of contribution value from an additional $10,000 of spend. What is its incremental ROI?

A. 2%

B. 12%

C. 20%

D. 120%

Correct Answer: C. 20%

Question 5

Which statement about MAPE is correct?

A. It is always suitable when actual values are zero

B. It is expressed in monetary units

C. It can be unstable when actual values are zero or close to zero

D. It measures marginal ROI directly

Correct Answer: C. It can be unstable when actual values are zero or close to zero

Question 6

What is the primary purpose of Monte Carlo simulation in campaign planning?

A. To remove all uncertainty from a forecast

B. To generate a distribution of possible outcomes under uncertain inputs

C. To guarantee causal attribution

D. To replace business constraints

Correct Answer: B. To generate a distribution of possible outcomes under uncertain inputs

Question 7

A shadow price associated with a budget constraint most directly estimates:

A. The historical cost per click

B. The value of relaxing the constraint by one unit

C. The total campaign revenue

D. The frequency cap

Correct Answer: B. The value of relaxing the constraint by one unit

Question 8

If a campaign has a pacing ratio greater than 1, it has:

A. Spent less than its planned cumulative amount

B. Exactly matched its planned cumulative amount

C. Spent more than its planned cumulative amount

D. Generated more incremental sales than forecast

Correct Answer: C. Spent more than its planned cumulative amount

Question 9

Why should video not automatically be eliminated solely because its platform-attributed ROI is lower than search’s?

A. Video always has zero cost

B. Video may generate incremental reach or effects not captured by the attribution rule

C. Search cannot generate conversions

D. Attributed ROI is always irrelevant

Correct Answer: B. Video may generate incremental reach or effects not captured by the attribution rule

Question 10

Which practice best supports responsible automated budget optimization?

A. Allowing unlimited automated spend changes

B. Using only platform-reported conversions

C. Maintaining documented assumptions, spending caps and human review

D. Ignoring privacy because optimization is mathematical

Correct Answer: C. Maintaining documented assumptions, spending caps and human review