Thomas Vladeck writes about Recast’s marketing mix modeling work, explaining how models can account for changing ad performance, channel reach, and uncertainty—and how marketers can use better data to make more reliable decisions.
They are most useful for technical thought leadership on marketing mix modeling methods, data quality, and vendor claims.
Best for
A marketing analytics firm seeking technical thought leadership on advanced marketing mix modeling.
Not for
A consumer brand seeking broad consumer reach or guaranteed distribution.
Key takeaways
When Meta and YouTube spend rise together, an MMM may not distinguish their effects.
Channel performance changes over time, so models need time structure rather than one constant.
Saturation ceilings should grow with a channel's reach, not penalize it for entering bigger markets.
Check first: Ask about audience roles, engagement, sponsorship formats, inquiry route, and terms before planning a paid placement.
Main ideas
When two models disagree about a channel, the problem is usually the data, not the algorithm · Channel performance changes over time, so models need time structure rather than a single constant · Saturation ceilings should grow with a channel’s reach, not penalize it for entering a bigger market
Thomas Vladeck writes about Recast’s marketing mix modeling (MMM), methods for estimating how marketing affects outcomes such as revenue.
Their posts explain why models can disagree, how channel performance and reach change over time, and how Recast’s research team develops and improves its models.
The examples show why they focus not only on model outputs, but also on whether the available data can support a confident marketing decision.
Vladeck makes technical modeling problems concrete by connecting them to decisions marketers face.
When Meta and YouTube spend rise together, they explain, an MMM may be unable to distinguish their effects.
The remedy is more variation in the data, including deliberate spend changes or holdout tests.
They also show how expanding channel reach can distort estimates when a model treats its old saturation limit as fixed, and how Recast’s Channel Footprint Scaling uses coverage history to address the problem.
Across these examples, Vladeck treats uncertainty and changing performance as features to model, not nuisances to hide.
A post on speeding up Bayesian models adds an engineering perspective: Recast reduced median fitting time without changing the model’s parameters and complexity.
Together, the four posts help readers understand what an MMM can and cannot tell them, and how model design affects budget decisions.
Thomas Vladeck writes about how marketing mix models are built and where they mislead.
Read more
Across the posts, they start with a modeling choice most tools make silently — constant ROI, a fixed saturation ceiling, independent channel effects or point estimates without uncertainty — and show what that choice costs marketers making budget decisions. Vladeck’s position is that most model disagreement comes from data and assumptions rather than algorithms; marketing performance genuinely changes over time and across geography; and the complexity of Recast’s Bayesian models is a deliberate tradeoff, not an accident.
When two models disagree about a channel, the problem is usually the data, not the algorithm
Vladeck argues that models often cannot separate channel effects because the spend data does not contain enough independent variation.
Read more
When Meta and YouTube scale at the same time, or a brand scales all paid channels together, the math fits "Meta did everything," "YouTube did everything," and "both contributed equally" equally well.
Vladeck calls this unidentifiability and treats the model's answer as correct behavior: it is reporting that the data does not support a confident claim.
The fix Vladeck recommends is not a better algorithm but creating variation — changing spend deliberately, running holdout tests, and making other planned changes to the marketing program.
When Meta and YouTube scale spend at the same time, the model cannot rule out competing explanations, making it hard to identify which channel drove incremental revenue lift.
Channel performance changes over time, so models need time structure rather than a single constant
Vladeck argues that most MMMs assume channel performance stays constant across the modeling period, which doesn’t match how paid media behaves.
Read more
Recast encodes two assumptions using Gaussian processes with a locally periodic kernel: days close together should perform similarly, as should days exactly one year apart, while far-apart days can differ more.
The intuition is that time has structure.
They say this lets seasonality emerge naturally and keeps results stable week over week.
Without it, a model might report that a top channel suddenly became the worst overnight, making confident decisions difficult.
If Meta performed well yesterday, the model expects similar performance today rather than a random swing in the opposite direction. Without this structure, the top channel could become the worst with no explanation.
Saturation ceilings should grow with a channel’s reach, not penalize it for entering a bigger market
Vladeck identifies a specific failure mode: a paid media channel that once reached 30% of a country now targets 80%, and both sales and spend climb.
Read more
A model that overweights past data treats the new spend as climbing an old, low-ceiling response curve and under-reports recent performance.
One that overweights recent data overwrites the old curve and over-reports past performance.
Either mistake can lead a brand to pull back spend on a channel that’s working as intended.
Recast’s Channel Footprint Scaling uses a channel’s coverage history so the saturation ceiling grows with its footprint.
A channel expands from 30% to 80% of national reach. Without coverage history, the model reads rising spend along an outdated saturation curve as the channel running out of room.
Recast’s models are slow on purpose; making them fast was an engineering choice, not a simplification
Vladeck argues against fitting a simple regression because it would assert things about marketing that Recast does not believe: that a channel’s ROI is exactly the same today as in 2024, that one number captures ad decay and saturation, or that point estimates without uncertainty are enough to decide with.
Read more
In Recast’s model, ROIs move over time, each paid channel gets its own saturation curve, holidays and promos take their own shapes, and every estimate carries uncertainty.
A 20-channel model carries roughly 30,000 parameters and evaluates itself against the data more than 250,000 times.
The research team then made it 5–20x faster without simplifying the model, loosening priors or dropping parameters.
Vikadeck frames the payoff as iteration speed: rerunning when something looks off, testing a different assumption before committing, and eventually letting customers change their own assumptions and watch forecasts update in near real time.
Median run time fell from several hours to about 14 minutes. The team replaced Stan’s general-purpose gradient machinery with hand-derived gradients for Recast’s exact model, fixed an inefficient memory layout, ran independent calculations concurrently and stopped tens of thousands of intermediate values from being written to disk when nothing downstream read them.
The complaint that Recast was “slow” becomes a design argument: Vladeck sees the long fit time as evidence that the model reflects how marketing works, and the speedup as restoring the practical benefits of iteration rather than correcting an error.
Answers the question: why do marketing mix models disagree even when the method seems sound?
Read more
Newcomers should start here: it frames model disagreement as a data-variation problem, not just an algorithm problem.
Vladeck explains unidentifiability: when channel spend moves in lockstep, such as Meta and YouTube scaling at the same time, the model cannot tell which channel drove incremental revenue.
The math can fit “Meta did everything,” “YouTube did everything,” or “both contributed equally.”
The model is not failing; it is showing that the data lacks enough variation to separate the causes.
The fix is to create variation through deliberate spend changes, holdout tests, and other marketing program changes.
For brands planning a campaign
For a campaign or measurement brief, use this post to ask whether the brand’s historical data can support the requested attribution claim. If spend is highly correlated, recommend holdout tests or planned variation rather than treating the model output as settled.
Answers: how can an MMM reflect performance that shifts week to week without producing unstable results?
Read more
Start here for a clear example of how Recast encodes time structure.
Most MMMs assume channel performance stays constant over the modeling period, but paid media does not work that way.
Vladeck describes using Gaussian processes with a locally periodic kernel: days close together should have similar performance, as should days one year apart, while far-apart days can differ more.
This lets seasonality emerge naturally and avoids a model claiming a top channel suddenly became the worst overnight without explanation.
For brands planning a campaign
Use this when evaluating whether a creator can explain time-varying measurement in plain terms. It supports asking for models that produce stable, believable week-over-week channel performance instead of noisy swings that cannot guide budget decisions.
Answers: what happens to an MMM when a paid channel reaches much more of the market than before?
Read more
It’s a concrete example of a modeling change tied to a real planning problem.
Vladeck describes a case where a channel once reached 30% of the country and now targets 80%.
A model that overweights past data may treat new spend as climbing an old, low-ceiling response curve and under-report recent performance; one that overweights recent data may overwrite the old curve and over-report past performance.
Recast’s Channel Footprint Scaling takes in coverage history so the saturation ceiling grows with the channel’s footprint instead of penalizing spend into a bigger market.
For brands planning a campaign
For brands expanding targeting, use this to ask whether the measurement approach accounts for changing addressable reach. It turns a modeling feature into a practical question: will the model penalize spend for entering a larger market?
Answers: why does a serious MMM take time to fit, and what changed?
Read more
Start here for a specific behind-the-scenes account of model complexity and engineering tradeoffs.
Vladeck says Recast does not fit a simple regression because that would assume channel ROI is constant, one number captures decay and saturation, and point estimates without uncertainty are enough.
A 20-channel Recast model carries roughly 30,000 parameters and evaluates against data over 250,000 times.
The research team made it 5–20x faster—median run time around 14 minutes, down from several hours—without simplifying the model, loosening priors, or dropping parameters.
The speedups came from hand-derived gradients, a better memory layout, parallelizing independent calculations, and removing intermediate values written to disk that nothing read.
For brands planning a campaign
Use this when assessing technical depth or asking about iteration speed. Faster fitting lets teams rerun a model when something looks off, test assumptions before committing, and eventually let customers change assumptions and see forecasts update near real time.
A recurring approach: model assumptions first, decision consequences second4
Unidentifiable MMM results when spend patterns move together1
Channel Footprint Scaling for channels that expand reach1
Time-varying channel performance and seasonality1
Faster Bayesian MMM without simplifying the model1
Unidentifiable MMM results when spend patterns move together
Vladeck explains why two marketing mix models can give completely different answers about the same channel: often, the issue is the data, not the model.
Read more
When Meta and YouTube scale spend at the same time, the model has trouble separating which one drove incremental revenue.
The same math can support “Meta did everything,” “YouTube did everything,” or “both contributed equally.”
They call this unidentifiability, common when predictors move in lockstep—for example, when spending the same amount on two channels day after day or scaling all paid media together.
Frames a statistical diagnosis with a plain-language example: the model is behaving correctly when it reports that the data cannot answer the question.
For brands planning a campaign
A useful angle for brands that treat MMM output as definitive: assess whether the spend data contains enough independent variation to support channel-level claims, and discuss holdout tests or deliberate spend changes as ways to create it.
Channel Footprint Scaling for channels that expand reach
Vladeck introduces Channel Footprint Scaling, a Recast research feature for cases where a paid media channel that once reached 30% of a country now targets 80%.
Read more
Without it, a model that overweights past data may treat new spend as climbing an old, low-ceiling response curve and under-report recent performance; one that overweights recent data may overwrite the old curve and over-report past performance. Recast can incorporate a channel’s coverage history so the saturation ceiling grows with its footprint instead of penalizing spend into a bigger market.
For brands planning a campaign
Relevant for brands whose media plans change geographic or audience reach: the topic supports a campaign about avoiding outdated saturation assumptions and asking vendors how coverage changes enter the model.
Vladeck argues that most MMMs assume channel performance stays constant throughout the modeling period, which does not match how paid media works.
Read more
Recast uses Gaussian processes with a locally periodic kernel to encode two assumptions: days close together should have similar performance, and days exactly one year apart should also have similar performance. This lets seasonality emerge naturally and helps prevent a top channel from suddenly appearing to become the worst overnight without explanation.
Explains the kernel through intuition—days close together, days a year apart, days far apart—before naming the technical term.
For brands planning a campaign
Useful for explaining why week-over-week stability matters in MMM decisions and evaluating whether a model captures time structure rather than treating every period as identical.
Vladeck explains why Recast models have taken a while to fit: they do not use a simple regression that assumes a channel’s ROI is exactly the same today as in 2024, that one number can capture how fast an ad decays or where a channel saturates, or that point estimates without uncertainty are enough to make decisions.
Read more
A 20-channel model carries roughly 30,000 parameters and evaluates itself against the data more than 250,000 times. The research team then made Recast 5–20x faster, with median run time around 14 minutes instead of several hours, by replacing general-purpose gradients with hand-derived ones, fixing memory layout, running independent calculations concurrently, and preventing unread intermediate values from being written to disk.
For brands planning a campaign
A technical differentiator: speed matters because analysts can rerun a model when something looks off, test a different assumption before committing, and eventually let customers change assumptions and watch forecasts and plans update in near real time.
A recurring approach: model assumptions first, decision consequences second
Across the posts, Vladeck starts with a modeling choice or limitation and shows what it means for a marketing decision: unidentifiable channels, saturation ceilings that lag behind coverage, unstable week-to-week performance, and slow iteration.
Read more
The posts pair statistical concepts with concrete numbers—30% to 80% coverage, 30,000 parameters, 250,000 evaluations, 14 minutes—without presenting the model as a black box.
For brands planning a campaign
For a thought-leadership assessment, the pattern is useful: Vladeck’s material is strongest when a brand needs a technical concept translated into a reason to change data collection, model setup, or vendor questions.
Thomas Vladeck writes as an insider at Recast, the marketing mix modeling company whose research and customers feature across the posts.
Read more
They describe Recast’s model development in the first person, and one post points customers to their Recast MDS team. The posts don’t mention sponsorships, affiliate arrangements, client work, or investments; ask about any relevant relationships before a campaign.
Recast
Employment
Vladeck writes in the first person about Recast's modeling and research work, describing what "our research team" shipped, how "our model" handles time-varying performance, and what changes mean for "our customers."
Read more
One post directs interested readers to their Recast MDS team. The posts indicate they work at Recast; they do not state a job title or start date.
Thomas Vladeck’s posts offer a clear starting point for understanding what a marketing mix model is doing when it gives you a number.
Read more
Read the four pieces here in order.
Start with the one on unidentifiability, which explains why two models can give opposite answers about the same channel and why the fix is usually more variation in the spend data, not a cleverer algorithm.
Then read the post on Channel Footprint Scaling, about a channel whose reach expands and an old saturation ceiling that no longer fits.
Next, the post on time-varying performance shows why days close together and days a year apart get similar weight in Recast’s models.
The final post covers model speed: they explain why complexity is a deliberate design choice and what the research team removed to make runs finish in minutes.
Across all four, the pattern is consistent: name an assumption a model makes silently, then show what it costs the person deciding where the next dollar goes.
If you follow measurement debates, these posts give you the vocabulary to ask vendors how their model handles moving ROIs, changing coverage and uncertainty — and to distinguish a real modeling tradeoff from a performance claim.
Sandoval writes about the same core question as Vladeck: whether a platform’s reported numbers reflect real incremental impact.
Read more
Her primer on incrementality covers holdout tests, geo-split experiments and modeled attribution, treating marketing mix models as one of several tools for getting past broken attribution. Both write for practitioners who have to justify spend without trusting platform dashboards.
Sandoval stays at the practitioner level, with short crash courses and regular roundups of paid media platform changes. Vladeck explains how the model is built — parameters, kernels, priors, gradients — from inside a modeling team rather than as an outside advisor.
Trippe also works at the seam between measurement and media execution.
Read more
His argument that a missed forecast is often a spend-adherence problem, not a model failure, aligns with Vladeck’s insistence that the data fed to a model determines what it can conclude. Both tell marketers to check what actually went out the door before blaming the model.
Trippe focuses on operations — reconciling partner spend against plan, owning cross-vendor accountability and timing Black Friday spend by channel lag — while Vladeck focuses on the model’s statistical structure. Trippe writes about running the media plan; Vladeck writes about modeling it.
Scogin writes detailed, practical LinkedIn teaching posts for the same audience: DTC marketing leaders trying to get more revenue from limited budget and headcount.
Read more
Both build arguments from specific operational experience and end with something readers can act on.
Scogin covers organizational and creative questions: how to staff a marketing function, and why more ad variations won’t fix a creative strategy that keeps making the same argument.
Vladeck covers measurement: how models separate channel effects and where their assumptions break.
Read Scogin for team and creative decisions, and Vladeck for how the numbers behind those decisions are produced.
Vladeck's published audience reads as a specialist marketing-analytics audience, making fit more relevant than scale.
Read more
The account has 7,376 LinkedIn followers as of 8 October 2026.
Posts cover marketing mix modeling, Bayesian modeling, channel saturation, and experimental design at a level that presumes familiarity with modeling practice.
That register suits a marketing analytics firm positioning thought leadership for practitioners and analytics buyers, but is unlikely to produce large consumer-style reach.
Before committing, ask whether they have audience survey results or another documented source describing followers' job function, seniority, and geography, and request recent engagement results for posts of comparable technical depth and format.
The material reviewed for Thomas Vladeck documents no sponsorship or partnership format.
Read more
The posts are technical and educational, covering marketing mix modeling choices, how Recast’s models are built and sped up, and how data structure affects what a model can identify.
They describe no creator sponsorship offering, deliverable, rate, exclusivity term, or booking process.
The only contact route mentioned is for existing customers to reach their account team to turn on a product feature.
The posts are written from inside Recast’s modeling and research team, so a commercial conversation would likely run through the company, not a personal creator partnership.
Before planning a sponsored engagement, ask Recast who handles partnerships and confirm the available formats, pricing, and availability.
Good fit: a marketing analytics firm wants technical thought leadership on advanced marketing mix modeling. Vladeck’s posts turn modeling decisions into practical questions for analytics teams. A post on unidentifiability explains why a model cannot separate Meta’s and YouTube’s contributions when both channels scale together—and why deliberate spend variation or holdout tests can help. Other posts connect changing channel reach, seasonality, and model uncertainty to measurement decisions. The material is a strong editorial fit for an audience evaluating MMM methods, data quality, or vendor claims. The content suggests a specialist audience, but does not identify job roles or report engagement. Before planning a paid placement, ask Vladeck or Recast about audience roles, engagement, available sponsorship formats, an inquiry route, and terms.
When to choose a specialist
Read more
Poor fit: a consumer brand wants a broad-reach lifestyle endorsement with a ready-to-book creator package. Vladeck’s posts focus on technical marketing measurement and Recast research, not general consumer products. They give a follower count but no demographic or engagement data to establish broad consumer reach. A buyer seeking a confirmed sponsored post, consumer audience, or guaranteed distribution should choose a creator whose relevant reach and commercial terms are available. Ask Vladeck whether they accept sponsorships and, if so, about formats, rates, booking process, and audience data.
A worked campaign example
Hypothetical brief: A marketing analytics firm wants to help analytics leaders recognize when channel-level MMM results are not supported by the data. Fit decision: Strong editorial fit, subject to confirming participation and commercial terms. Campaign angle: “Before debating which channel won, ask whether the data can tell.” Build an educational post or discussion around Vladeck’s example of Meta and YouTube scaling together: when spend moves in lockstep, multiple explanations can fit the same outcome. Use that example to prompt teams to check for independent variation and consider holdouts or deliberate spend changes before presenting channel-level conclusions as settled. Next step: Ask whether Vladeck would participate in an educational collaboration, what format and review boundaries would apply, and whether recent engagement results are available for similarly technical posts. Confirm those points before promising a sponsored deliverable or audience outcome.
Choose Gavin Trippe when an analytics firm's clients question model accuracy and the first step is checking whether media partners spent the plan.
Read more
Trippe recommends comparing each partner's actual spend with plan before blaming the model, and treating a forecast miss as a possible adherence problem.
That practical framing suits firms helping clients distinguish execution failures from modeling failures.
Trippe also offers concrete channel timing guidance, such as starting long-lag TV and audio well before Black Friday so purchases land during the sale window.
Trippe covers media operations and planning, not model specification choices such as unidentifiability, time-varying performance, Bayesian fitting speed, or channel footprint scaling. Buyers seeking advanced marketing mix modeling thought leadership will find less technical depth than in Vladeck's work.
Choose Cynthia Sandoval when the audience is earlier in measurement maturity and needs a plain-language reason to move beyond platform-reported ROAS.
Read more
Her incrementality post explains holdout tests, geo-split experiments, and modeled attribution as practical responses to broken attribution—a useful starting point for media buyers, CMOs, and founders who are not yet asking about model parameters. Her platform-change roundup suits campaigns needing a quick digest of current paid media shifts across Meta, Google, YouTube, privacy, and AI tools.
Sandoval's incrementality explainer offers a clear conceptual overview but little detail on MMM model design or the parameter-level decisions Vladeck describes.
The platform-changes post is a news roundup, useful for currency rather than durable methodology.
Analytics firms seeking advanced MMM thought leadership will find less modeling detail than in Vladeck's work.
Choose Sean Scogin when the campaign goal is DTC growth execution rather than measurement methodology.
Read more
Their marketing-function post lays out a concrete hiring and resource plan for a low eight-figure brand, while their creative-strategy post argues that more ad variations will not fix rising Meta CPA if every ad makes the same argument.
For buyers helping brands interview customers, mine reviews, and build genuinely different creative arguments, Scogin's work is a better fit than MMM model internals.
They also offer to audit a brand's reviews and Meta ad library and return one underused customer segment.
Scogin focuses on team building, creative strategy, customer research, and agency services—not advanced MMM. The work relies more on customer research and creative testing than model-based measurement, making it useful for DTC operators but less suited to questions about unidentifiability, seasonality modeling, or Bayesian fitting speed.
This page covers four LinkedIn posts written by Thomas Vladeck about marketing mix modeling and Recast's research: why two models can reach opposite conclusions about the same channel, how a channel's expanding reach breaks a fixed saturation ceiling, how Recast handles performance that changes over time, and how the research team cut median model fitting time from several hours to about 14 minutes.
The posts have no publication dates and include no images or video transcripts, so their titles or snippets alone do not convey their full content.
These posts show detailed, example-driven explanations of modeling choices and their consequences for budget decisions; they do not establish a wider career history or signature books, shows, or earlier work, and no audience composition or engagement figures are stated.
Statements about employment and company work reflect what the posts themselves say.
Where other creators appear for comparison, their contributions come from posts on their own accounts and are offered as reading context, not endorsement.
