Frameworks & methodology for marketing budget allocation & forecasting

A fixed pot, eleven channel lines, and forty minutes with the CFO to justify why line three doubles while line seven halves. That is the decision this lesson arms you for. The arguments that survive that meeting are marginal ones: not what a channel returned on average last year, but what the next $100,000 into it returns compared with the next $100,000 anywhere else. Average ROIROIReturn on Investment: the ratio of net profit to the cost of an investment. A 300% ROI means each dollar invested returns $3.View full definition → ranks channels. Marginal ROI moves money, and every method below is a way of estimating it.

Core concept: marginal ROI is the unit of allocation

Take two channels. Paid search returns $6 per dollar on average, affiliate $3. The obvious move is to fund search. Now add the shape: search is bidding against itself on branded terms, the last $2 million of it returned about $1.40 per dollar, and affiliate is small enough that the next $2 million would still hold near $3. The reallocation runs the other way. The average told you the opposite of the right answer.

Every allocation method worth running estimates that slope: the change in revenue for the next unit of spend, per channel, at the level you are spending today. Two things follow. Allocation is only ever locally correct, so a plan built on last year's slopes decays as auction prices, competitors and creative change. And you can only fit curves to the portion of spend that actually buys attention: agency retainers and martechmartechThe connected set of software tools a marketing team uses to plan, run, measure and automate campaigns across channels.View full definition → licences do not sit on a response curve, so the split the foundations lesson draws is what feeds the model.

Key sub-concept 1: the models that produce a curve

Marketing mix modelling (MMMMMMA statistical approach that estimates how each marketing channel and other factors drive sales, guiding budget allocation.View full definition →) regresses sales against spend by channel over time, controlling for seasonality, price, distribution and competitor activity, and returns a coefficient and a saturation shape for each line. It sees offline media, it survives cookie deprecation, and it needs two to three years of genuine variation in spend to say anything useful. A vendor build runs in the low hundreds of thousands of dollars a year; open-source implementations exist, but the cost moves to your data team rather than disappearing.

Multi-touch attributionMulti-touch attributionA method that distributes conversion credit across all marketing touchpoints in the customer journey, rather than crediting only the first or last interaction.View full definition → works from the other end: individual journeys, credit distributed across trackable touchpoints, fast enough to act on weekly. It cannot see television, sponsorship, out-of-home or word of mouth, so anything it cannot track scores zero by construction. Set an annual split from MTA alone and you will systematically defund the media it is blind to.

Nielsen, which sells mix modelling and therefore has an interest in the answer, has reported that a substantial share of brands sit below their optimal spend level, where additional media would still clear breakeven. Hold onto that, because an optimiser constrained to a fixed pot can never return the recommendation "spend more". It will shuffle the same dollars indefinitely and call the result efficiency.

Key sub-concept 2: response curves and the extrapolation trap

Response curves plot spend against output and flatten as audiences saturate, frequency stacks on the same people and auction prices rise to meet your bid. Your job is to spend up to the point where each channel's marginal return equals every other channel's, then stop.

The trap is that a curve is only fitted where you have data. If a channel has run between $1 million and $3 million for three years, the model's opinion about $8 million is an assumption dressed as a finding, and the curve almost always overstates what happens out there. The same applies at the bottom end: if you have never taken a channel to zero, the model cannot separate its contribution from the baseline the foundations lesson describes. This is why serious teams pair MMM with geogeoThe practice of making your brand and content visible and citable inside AI-generated answers from tools like ChatGPT, Gemini and Perplexity.View full definition → holdouts, switching a channel off in matched regions for six to eight weeks and using the observed gap to calibrate the curve rather than trusting the regression alone. Budget roughly 1 to 2 percent of annual spend for that testing. It is the cheapest insurance in the plan.

Key sub-concept 3: the brand versus performance split

In 2019 Adidas' global media director Simon Peel said publicly that the company had over-invested in performance and digital at the expense of brand. The working split had been roughly 77 percent performance to 23 percent brand, while Adidas' own econometric modelling attributed around 65 percent of revenue to brand activity. The mechanism was ordinary: e-commerce attributionattributionA framework for assigning credit to the touchpoints that contributed to a conversion, so you can measure which channels and interactions actually drive results.View full definition → credited paid search and last-click channels with demand that broadcast and sponsorship had created, and every quarterly review reinforced the error.

That is the second-order consequence to watch. Performance channels harvest demand, so cutting brand makes performance metrics look better for two or three quarters before the demand pool drains. The reporting improves while the business weakens, which is exactly the pattern that persuades a board to cut again.

Binet and Field's 60:40 brand-to-activation ratio is a reasonable prior, not a law. It moves with purchase cycle, margin and category growth: a subscription business with monthly repurchase and a durable-goods brand bought once a decade do not share a split. Treat 60:40 as a sanity check on the number your curves produce, and if your model says 85 percent performance, ask what it cannot see before you believe it.

Key sub-concept 4: zero-based rebuilds

A zero-based rebuild starts every line at nothing and requires a projected return to earn its allocation, instead of indexing off last year. Unilever ran this across marketing from 2016, reporting roughly €500 million of savings in 2017 by making about 30 percent fewer ads and consolidating agency and production costs, without a corresponding drop in media weight.

The failure mode is running it on everything, every year. Zero-basing rewards whatever is easiest to evidence, which means performance channels with clean tracking survive and brand and research lines die, and after two cycles you have a cheaper marketing function with a weaker asset. Rotate it: one category, one agency contract or one channel cluster per year, with a floor under brand spend that the rebuild is not allowed to touch.

Rolling forecast mechanics

An annual budget locked in October is wrong by February. A rolling forecast re-cuts the next 12 to 18 months at each quarter close, carrying the plan forward rather than defending the original number.

The mechanics that make it work:

• Hold 10 to 15 percent of working spend unallocated at the start of the year, as a reallocation reserve rather than a contingency. Contingency gets swept by finance; a reserve with a documented decision date does not.
• Refit curves at each close using the quarter's actuals, and version the assumption list. When the forecast changes, the board should be able to see which assumption moved, not just which number did.
• Set a variance trigger. Only move money when the modelled gap in marginal return between two channels exceeds the model's own confidence interval. Below that, you are trading noise and paying rebooking costs for the privilege.
• Forecast in three cases with stated assumptions, and state the lag: most brand and B2B activity shows revenue impact 60 to 180 days out, so a Q4 reallocation mostly changes Q1 and Q2 results.

The constraint nobody models is cancellability. Upfront television, sponsorships and out-of-home are committed months ahead with narrow cancellation windows, while search and social can be changed in a fortnight. So a mid-quarter re-cut lands almost entirely on the flexible lines, which are usually the performance lines, and repeated re-cutting quietly shortens your average time horizon. Keep the reserve inside the flexible lines deliberately, treat committed media as fixed until the next planning window, and count how many times a year you actually pull the lever. More than four is churn.

CMO action items

• Commission a mix model if annual working spend exceeds $10 million. At vendor prices in the low hundreds of thousands, it pays back inside one planning cycle if you act on the reallocation, and not at all if you file it.
• Build response curves for your four largest channels before the annual cycle, then calibrate at least one of them with a geo holdout. An uncalibrated curve is a hypothesis with a chart attached.
• Zero-base a rotating 20 percent of the budget each year, with an explicit floor under brand spend so the exercise cannot become a slow defund.
• Write the reallocation trigger down before the year starts: reserve size, review dates, and the minimum marginal-return gap that justifies moving money.

Common mistakes that kill results

• Reading attribution output as ground truth. Last-click and most MTA models over-credit branded search and retargetingretargetingShowing ads to users who have previously visited your site or interacted with your brand, to bring them back and drive conversion.View full definition →, the two channels that sit closest to a purchase already decided. Adidas' 2019 admission is the clean version of this: the measurement system pointed at the harvesting channels, the econometrics pointed at brand, and the money followed the wrong one for years.
• Forecasting from averages. A single-point forecast ("12,000 leads next quarter at this spend") offers false precision. Scenario ranges with named assumptions survive contact with a market shift; point forecasts cost you credibility the first quarter they miss.
• Judging campaigns at 30 days. If brand and B2B response arrives over 60 to 180 days, a 30-day ROI read systematically under-rates the channels with the longest payback, and you reallocate away from them every cycle.
• Optimising a fixed pot when the honest answer is a bigger one. If the marginal return on your best channel is still well above your cost of capitalcost of capitalThe blended rate a company pays to finance itself through debt and equity. It sets the minimum return an investment must clear to create value.View full definition → at the current ceiling, the recommendation is more budget, and no reallocation model will ever tell you that on its own.