Marketing Value at Risk: How Much Can Your Budget Lose?

Abstract

Marketers forecast expected returns but never quantify how much they stand to lose. We adapted Value at Risk (VaR) -- the standard downside risk metric in finance -- to marketing portfolios using data from 1,221 Marketing Mix Models across 123 brands, 22 markets, 83 channels, and $1.85B in measured ad spend (2020-2025). The result: a single number that answers "with 95% confidence, what is the maximum revenue shortfall from our marketing portfolio in any given quarter?" Across our dataset, the median brand carries a quarterly Marketing VaR of 22-31% of expected marketing-driven revenue -- meaning in 1 out of 20 quarters, the portfolio will underdeliver by at least that amount. Most CMOs have no idea this exposure exists. This is the missing metric that turns marketing from gambling into capital management.

Keywords: marketing value at risk, VaR, marketing risk measurement, budget risk, downside risk, Bayesian MMM, portfolio management.

Table of Contents

• Abstract

• The Metric Marketing Does Not Have

• What Value at Risk Actually Is

• Adapting VaR to Marketing Portfolios

• Computing Marketing VaR From Bayesian MMMs

• What the Data Shows: VaR by Channel

• Portfolio VaR: The Diversification Effect

• A Real Case: Detecting Hidden Downside Before It Materialized

• How to Run This Analysis With Your Data

• Known Limitations

The Metric Marketing Does Not Have

Every CFO managing a financial portfolio knows two numbers: expected return and Value at Risk. The first tells you what you expect to make. The second tells you the worst-case scenario at a defined confidence level.

Marketing has the first number. Every quarterly plan includes a revenue forecast tied to ad spend. "We expect to generate $12M in marketing-driven revenue from a $3M budget." The second number does not exist.

Ask a CMO: "What is the maximum you could lose on that $3M budget?" You will get silence, or hand-waving, or a deflection to ROAS benchmarks. Nobody has a quantified answer because nobody computes it.

This is not a trivial gap. In finance, a portfolio manager who could quote expected return but not VaR would be fired before lunch. They would be flying blind on exactly the dimension that determines survival -- the downside.

Marketing budgets are the same category of capital allocation. They carry the same structural risks: channel performance fluctuates, auction dynamics shift, creative fatigue degrades returns, platform algorithm changes redistribute outcomes. Yet the industry has no standard metric for quantifying how bad a bad quarter can get.

We built one. Using the posterior distributions from 1,221 Bayesian Marketing Mix Models, we computed Marketing VaR across our entire client base. The results reveal that most brands carry 2-3x more downside exposure than their planning assumptions imply.

What Value at Risk Actually Is

Value at Risk was developed by JP Morgan in the early 1990s to answer a deceptively simple question: "How much could we lose tomorrow?"

The formal definition:

VaR(alpha) = the maximum loss that will not be exceeded with probability (1 - alpha) over a given time horizon.

In practice, a 95% daily VaR of $10M means: "On 95% of days, we will not lose more than $10M. On 5% of days, we might."

Three components define any VaR calculation:

VaR became the global standard because it compresses complex portfolio risk into a single, communicable number. Regulators require it. Boards review it. It forces institutions to confront their actual exposure rather than hiding behind expected returns.

Marketing needs the same discipline.

Adapting VaR to Marketing Portfolios

The translation from financial VaR to marketing VaR requires redefining what "loss" means in a marketing context.

In finance, loss = decline in portfolio value. In marketing, loss = shortfall in expected marketing-driven revenue.

Here is the formal definition we use:

Marketing VaR(95%, quarterly) = the maximum revenue shortfall relative to the expected marketing-driven revenue, such that there is only a 5% probability of the actual shortfall exceeding this amount in any given quarter.

In plain language: if your plan says marketing will generate $10M in revenue this quarter, and your Marketing VaR is 25%, then there is a 5% chance that marketing will underdeliver by $2.5M or more. Your worst realistic quarter produces $7.5M or less in marketing-driven revenue.

The formula:

Marketing VaR = Expected Revenue - Revenue at 5th Percentile of the Posterior Distribution

Or expressed as a percentage:

Marketing VaR % = (Expected Revenue - P5 Revenue) / Expected Revenue

Where:

This is where the Bayesian MMM becomes essential. Traditional regression-based models produce point estimates. They tell you "Google ROAS is 4.37x" but give you no distribution around that number. You cannot compute VaR from a point estimate.

A Bayesian MMM -- the kind Cassandra runs using PyMC and our proprietary priors -- produces a full posterior distribution for every channel coefficient. Each coefficient represents thousands of plausible return values given the data. That distribution is the raw material for VaR calculation.

Computing Marketing VaR From Bayesian MMMs

The computation proceeds in four steps.

Step 1: Extract posterior distributions. For each channel in the model, the Bayesian MMM produces a posterior distribution of the incremental revenue per unit of spend. This is not a single number -- it is typically 4,000-10,000 draws from the posterior, each representing a plausible value given the observed data.Step 2: Simulate portfolio revenue. For a given budget allocation across channels, draw one value from each channel's posterior distribution and compute total marketing-driven revenue. Repeat 10,000 times. This is a Monte Carlo simulation that propagates channel-level uncertainty into portfolio-level uncertainty.Step 3: Construct the revenue distribution. The 10,000 simulated total revenue values form the portfolio revenue distribution. This distribution captures both individual channel risk and the correlations between channels (because the posterior draws preserve covariance structure).Step 4: Read off VaR. The 5th percentile of the revenue distribution is the VaR threshold. Everything below this point represents the tail risk that exceeds 95% confidence.

Marketing VaR (95%) = P50(Revenue Distribution) - P5(Revenue Distribution)

This approach is superior to parametric VaR (which assumes normal distributions) because marketing returns are demonstrably non-normal. They are skewed left -- downside surprises are larger and more frequent than upside surprises. The Monte Carlo method captures this asymmetry directly from the data.

What the Data Shows: VaR by Channel

We computed individual channel VaR across our dataset of 1,221 MMMs. The results reveal that channels with similar median ROAS carry dramatically different downside exposure.

Channel-Level VaR (95%, Quarterly)

The CI Width column comes directly from the Risk-Adjusted ROAS analysis we published previously. VaR extends that work by translating the width of the confidence interval into a dollar-denominated downside number.

Three findings stand out:

1. Meta has the lowest VaR among major paid channels (14%). Meta's narrow confidence interval (0.78) translates directly into a narrow downside tail. This does not mean Meta is the best channel. It means Meta is the most predictable. For budgets where consistency matters more than upside -- e.g., a brand with thin margins and fixed cost obligations -- Meta's low VaR is a genuine advantage. As we showed in the 2026 Media Effectiveness Benchmarks, Meta's consistency is its defining feature when measured properly.2. Amazon's VaR is nearly 3x Google's despite similar median ROAS. Amazon's 41% quarterly VaR means that in a bad quarter, more than two-fifths of the expected marketing-driven revenue from Amazon disappears. The CI width of 2.91 -- nearly as wide as the median ROAS itself -- tells you the outcomes are almost as likely to be near-zero as they are to be near the median. This is not a channel you build a budget plan around unless you have a high tolerance for variance.3. The gap between median ROAS and VaR is the information standard reporting hides. Google and Amazon have nearly identical median ROAS (4.37x vs. 3.12x). A standard ROAS report presents them as "both above 3x, both good." VaR reveals that Google's bad quarter costs you 18 cents on the dollar, while Amazon's bad quarter costs you 41 cents. These are not comparable investments.

Portfolio VaR: The Diversification Effect

Individual channel VaR overstates total portfolio risk. This is the same insight that underpins the Marketing Efficient Frontier: channels that are imperfectly correlated provide diversification, and the portfolio's downside is smaller than the weighted sum of individual downsides.

We computed portfolio VaR for a representative allocation: 30% Google, 30% Meta, 15% TikTok, 15% TV, 10% Amazon.

The portfolio VaR of 17.1% is materially lower than the 22.3% you would compute by simply averaging channel VaR. The reduction comes from imperfect correlation between channels: when Meta underperforms in a given quarter, Google does not necessarily underperform by the same amount.

This has a direct implication for budget concentration. A brand that allocates 80% of budget to a single channel -- even a low-VaR channel like Google (18%) -- has a portfolio VaR close to 18%. A brand that spreads allocation across five channels with moderate individual VaR can achieve a portfolio VaR of 14-17% through diversification alone.

We ran this analysis across allocation scenarios:

Portfolio VaR by Concentration Level

The optimized allocation -- derived from the efficient frontier -- achieves the lowest VaR not by avoiding high-risk channels entirely but by sizing positions according to their risk contribution and correlation structure. This is capital management applied to marketing.

A Real Case: Detecting Hidden Downside Before It Materialized

In Why Attribution Misleads Budget Decisions, we documented a case where shifting 25% of budget from Google Search to Google Video yielded +18% incremental conversions.

The VaR analysis reveals why this worked beyond the ROAS differential.

Before the shift, the brand's portfolio had a quarterly VaR of 28%. The portfolio was concentrated in Google Search, which attribution data rated at 255x more efficient than Video. But the MMM told a different story -- and the posterior distributions told even more.

Google Search's revenue posterior was wide. The 90% credible interval for CPA ranged from $270 to $632 -- a $362 spread. Google Video's posterior was tight: $110 to $150, a $40 spread. In VaR terms:

Search carried nearly 3x the downside risk of Video on a per-conversion basis.

After the 25% budget shift to Video:

• Expected conversions increased by 18%

• Portfolio VaR dropped from 28% to 19%

The brand did not just improve expected returns. It simultaneously reduced tail risk. This is the definition of a dominant move in portfolio theory: higher return and lower risk. Such moves are rare -- but they exist precisely when the prior allocation was made without risk information, which is the norm in attribution-driven planning.

How to Run This Analysis With Your Data

Step 1: Get a posterior distribution

You need a Bayesian MMM, not a frequentist one. The posterior distribution is the foundation of VaR. If your current MMM produces point estimates with standard errors, those can be approximated as normal distributions -- but this understates tail risk. A full Bayesian model (PyMC, Stan, or Meridian) gives you the actual posterior draws.

If you do not have a Bayesian MMM, you can approximate channel-level VaR using historical weekly ROAS data:

This approximation treats each week as an independent draw, which understates serial correlation. Use it as a lower bound.

Step 2: Simulate portfolio revenue

For each of 10,000 iterations:

• Draw one ROAS value per channel from its posterior (or from the empirical weekly distribution)

• Multiply by the planned quarterly spend per channel

• Sum across channels to get total simulated portfolio revenue

Store all 10,000 totals. This is your portfolio revenue distribution.

Step 3: Compute VaR

Marketing VaR ($) = Median(Revenue Distribution) - Percentile(Revenue Distribution, 5)

Marketing VaR (%) = VaR($) / Median(Revenue Distribution)

Step 4: Stress-test concentration risk

Re-run the simulation under different allocation scenarios:

• Current allocation

• Equal-weight across channels

• Remove the highest-VaR channel entirely

• Shift 10%, 20%, 30% from highest-VaR to lowest-VaR channel

Plot portfolio VaR against expected return for each scenario. You will see your own marketing efficient frontier emerge -- and you will see exactly where your current allocation sits relative to the frontier.

Step 5: Set a VaR budget

Define an acceptable Marketing VaR for your organization. This is a governance decision, not a technical one. A brand with 60% gross margins and a war chest can tolerate a 30% VaR. A brand operating at breakeven with debt covenants tied to revenue targets needs VaR below 15%.

Once your VaR budget is set, it becomes a constraint on allocation: no portfolio that exceeds the VaR limit is admissible, regardless of its expected return. This is how finance has operated since Basel II. Marketing can adopt the same discipline.

Known Limitations

Stationarity assumption. VaR assumes the distribution of future returns resembles the distribution of past returns. Marketing channels are non-stationary -- iOS 14.5 permanently altered Meta's return distribution; AI-generated content is changing creative dynamics in real time. VaR should be recomputed quarterly with recent data, not treated as a static number.Correlation instability. Channel correlations shift under stress. When the economy contracts, all paid channels tend to underperform simultaneously -- correlations spike toward 1.0, and the diversification benefit disappears exactly when you need it most. This is the same "correlation breakdown" problem that plagued financial VaR in 2008. Conditional VaR (CVaR), which measures the expected loss given that VaR is exceeded, is a useful complement.Model risk. VaR is only as good as the underlying MMM. If the model misattributes TV's halo effect to Google branded search, then Google's posterior distribution is too tight and its VaR is understated. Model validation -- through geo-experiments, holdout tests, and out-of-sample prediction -- remains essential, as we documented in Why Attribution Misleads Budget Decisions.Sample composition. Our dataset skews toward mid-to-large brands with $500K+ monthly ad spend. Smaller budgets may have wider posteriors (more uncertainty per channel) and higher VaR as a result.VaR does not capture maximum loss. VaR tells you the threshold of the 5% tail; it does not tell you the average loss within that tail. A brand with 20% VaR might lose 20% in a bad quarter or 50%. CVaR (expected shortfall) addresses this, and we plan to publish that analysis separately.Every channel estimate in Cassandra includes a full posterior distribution -- not a point estimate. That means Marketing VaR is a native output, computed automatically for your portfolio every time the model updates. If you want to know the actual downside exposure of your marketing budget, book a call.