Essay · 2026
The Incrementality Frontier
Portfolio selection for media capital — co-movement, causal return, and the gap between what platforms claim and what the mix actually creates.
Marketing · media mix · portfolio theory · incrementality
“Diversification is the only free lunch.” — attributed to Harry Markowitz1
Most media plans still treat the budget like a shopping list. Pick the platforms that performed last quarter, add a little for the new format everyone is talking about, and negotiate rates until the spreadsheet balances. That process optimizes channels against attributed scores. It rarely optimizes the portfolio for incremental outcomes.
Modern portfolio theory (MPT)2 was built for allocating capital when returns are uncertain and risks co-move. Media has the same structure—and an extra distortion: the numbers we optimize are often not the returns we get. Last-click ROAS, platform attribution, and multi-touch models can all reward channels for demand that would have arrived anyway. The real asset is not “claimed conversion.” It is incrementality: the outcome that would not have happened without the spend.
What follows is a way to hold both ideas at once. Distribute media capital like a careful investor—expected return, correlated risk, rebalancing—and move the objective of optimization from vanity efficiency to incremental contribution of the mix. The lab at the end lets you feel the gap between the two.
Expected Return
In finance, an asset’s expected return is not last month’s performance. It is a forward-looking estimate under uncertainty. In media, we often invert that: we treat yesterday’s ROAS or CPA as if it were a promise, then act surprised when the market shifts—or when a geo-holdout shows half the “performance” was organic demand wearing a paid label.
A portfolio view forces a better question: what outcome am I purchasing with this dollar of attention? Not “which platform had a good week,” but which combination of reach, mental availability, conversion, and learning best serves the period’s objectives—and with what confidence that the spend caused the movement.
That means separating short-cycle performance signals from brand and demand effects that show up late3, and treating both as components of expected return only to the extent they are incremental. A mix built only on attributed short-cycle scores will overweight channels that harvest existing intent and underweight the ones that create the next period’s cheaper demand.
Practically: write the objective first, then the few quantitative results that would prove progress. Only then allocate. If a channel cannot be connected to an incremental result you care about—only to a dashboard that assigns credit—it is inventory with a story attached, not yet an asset in the portfolio.
The vertical axis is not “reported ROAS.” It is expected incremental return. Points that look efficient on attributed scores often sit inside the frontier once you remove non-causal credit.
Correlated Risk
Portfolio risk is not the sum of individual risks. It depends on how assets co-move. Two volatile assets that spike and crash together protect you less than two with lower correlation—even if each looks “risky” alone.
Media has the same structure. Overweighting three social formats that share auction dynamics, attention trends, and creative fatigue is not diversification. It is concentration with multiple invoices. When the feed changes or CPMs reprice, the stack moves together—and so does any measurement error built into the same attribution system.
A portfolio-minded planner asks: where do my outcomes co-vary? Platform family, audience, creative system, seasonality, and measurement method all induce correlation. True diversification mixes mechanisms—intent, discovery, open-web reach, offline, CRM—not just line items inside one garden. It also diversifies how you learn causality: platform reporting alone is a single, highly correlated sensor.
From Optimization to Incrementality
Optimization is not the enemy. Optimizing the wrong objective is. When the objective function is platform-attributed conversion, the system will correctly maximize credit—including credit for buyers who would have converted without the ad. That is not a bug in the algorithm. It is a bug in the loss function.
Incrementality reframes the unit of return: the lift relative to a counterfactual world without the spend (or without that slice of the mix)4. Methods differ—geo experiments, PSA tests, conversion lift, synthetic controls, well-specified media-mix models—but the philosophy is shared: pay for what you cause, not for what you merely co-occur with.
Shifting optimization toward incrementality changes allocation in predictable ways:
Harvesting channels (strong intent, high claim, low lift) lose weight · Creating channels (weaker short-cycle scores, real demand generation) gain it when evidence supports them · Overlapping channels that double-count the same conversion stop looking additive · A deliberate learning sleeve becomes rational capital expenditure, not “waste”
Portfolio theory and incrementality reinforce each other. MPT asks you to care about joint outcomes under co-movement. Incrementality asks you to define those outcomes causally. Together they produce a sharper frontier: mixes that maximize incremental expected return per unit of risk—not mixes that maximize the sum of flattering dashboards.
The operational move is not “stop optimizing.” It is: change what you optimize for, measure the gap between attributed and incremental return, and rebalance when that gap—not last week’s ROAS—moves.
Same loop at any budget scale. Tools change; the objective function should not quietly revert to attributed ROAS.
Lab — Attributed versus incremental portfolio
The sketch below is a teaching model, not a media-mix engine. Three media vehicles in the same class—paid search, social, and display—share a budget. Each has a claimed efficiency (what a dashboard might report) and an incrementality rate (what share of that claimed return would not have happened without the spend). A global credit overlap control simulates how much those vehicles co-claim the same outcomes—raising attributed totals while leaving true incremental value behind.
Drag the controls. Watch the gap between the flattering sum and the incremental portfolio. That gap is where optimization lies to you if you never change the objective.
Budget weights re-normalize to 100%. Claimed efficiency is an attributed-style multiple on spend. Incrementality is the causal share of that claim. Overlap inflates co-claimed credit without creating new lift.
Preset values are pedagogical calibrations of patterns in the academic literature—not replications of any single study’s point estimates. Brand-intercept follows field evidence that branded paid search often claims demand that would have arrived anyway; observational overclaim follows comparisons of platform-style or matching methods vs randomized lift; experiment-informed mix follows the case for causal measurement and rebalancing budget across vehicles with higher proven lift. Drag sliders to leave a preset.
Teaching model. α and overlap are stylized from research patterns5–7; real programs estimate them with experiments and models. The lab makes the arithmetic of the attribution gap impossible to ignore.
Teams that already work with shared objectives and clear feedback loops will find this shift easier: the budget becomes another instrument with a clef (what “good” means), a tempo (when you rebalance), and room to improvise inside structure. Without incrementality in the objective, “optimization” is rearranging noise—and diversification is just more ways to co-claim the same conversion.
The free lunch, if there is one, is not a new platform. It is a mix whose risks do not all fail on the same day, scored on outcomes you actually caused—and a habit of rebalancing when evidence of lift, not fashion, moves the frontier.
- The “free lunch” line is widely attributed to Markowitz in practitioner discourse; the formal foundation is his work on portfolio selection and diversification. ↩
- Markowitz, H. (1952). Portfolio Selection. The Journal of Finance, 7(1), 77–91. ↩
- Binet, L., & Field, P. (2013). The Long and the Short of It. IPA. ↩
- On causal measurement in advertising, see industry practice around conversion lift and geo experiments; the lab is pedagogical, not a substitute for experimental design. ↩
- Brand-intercept search preset. Blake, T., Nosko, C., & Tadelis, S. (2015). Consumer Heterogeneity and Paid Search Effectiveness: A Large-Scale Field Experiment. Econometrica, 83(1), 155–174. Large-scale experiments at eBay found branded paid search largely intercepted navigational demand—attributed “performance” with little incremental purchase effect. The preset stylizes a paid-search-heavy mix of the three vehicles: high claimed efficiency and very low α on paid search, with elevated co-claim across vehicles. ↩
- Observational overclaim preset. Gordon, B. R., Zettelmeyer, F., Bhargava, N., & Chapsky, D. (2019). A Comparison of Approaches to Advertising Measurement: Evidence from Big Field Experiments at Facebook. Marketing Science, 38(2), 193–225; Lewis, R. A., & Rao, J. M. (2015). The Unfavorable Economics of Measuring the Returns to Advertising. Quarterly Journal of Economics, 130(4), 1941–1973. Observational and selection-prone methods often overstate lift relative to RCTs; power and endogeneity make “dashboard efficiency” a noisy proxy for causality. The preset uses a more even split across paid search, social, and display with moderate co-claim inflation. ↩
- Experiment-informed mix preset. Same experimental literature as fn. 5–6 on replacing attributed scores with lift estimates when reallocating across vehicles; plus Binet, L., & Field, P. (2013). The Long and the Short of It. IPA—on not starving longer-horizon investment that last-click under-credits (often carried more by display than by intercept search). The preset shifts weight toward higher-α vehicles and lower overlap. ↩