-1.png?width=1132&height=292&name=Asset%201@2x%20(12)-1.png "Odins.ai Logo")
How it works — Bayesian approach
Why Bayesian beats old-school regression — especially if you're not in America
Traditional marketing mix models need mountains of data to work. Bayesian models are smarter about it — they combine what you already know with what the data shows. Here's why that matters, and why it matters even more if you're not sitting on 50 state-level datasets.
Traditional regression: powerful, but hungry for data
Classic frequentist regression — the kind from every statistics textbook — works by finding patterns purely from data. More data, better patterns. Simple.
The problem: marketing data isn't simple. You've got maybe 150 weeks of history, 10–20 channels to estimate, seasonality muddying everything, and channels that interact with each other. A traditional model looks at that and says: 'I need more data.'
In the US, you can split the analysis across 50 states — suddenly you've got 50 time series instead of one, and the math works beautifully. In Norway? The UK? Australia? You've got one time series for the whole country. And that's where the old approach starts to creak.
Bayesian modeling: start with what you know, let the data refine it
Bayesian statistics flips the script. Instead of starting from zero, you start with reasonable assumptions — what statisticians call 'priors' — and let the data update them.
Here's the thing: you already know roughly how your marketing works. You know TV probably has a longer lag than search ads. You know there are diminishing returns. You know branded search captures existing demand more than it creates new demand.
A Bayesian model starts with those beliefs and systematically tests them against your actual results. Where the data agrees, confidence goes up. Where it disagrees, the model adjusts. You end up with honest estimates that work even when data isn't abundant — because the model isn't starting from nothing.
You don't need 50 states to get a useful model
In the US, companies like Google built their MMM tools (like Meridian) assuming you'll feed in data from 50 states or hundreds of DMAs. That's a luxury most of the world doesn't have.
If you're running marketing in Norway, the Netherlands, Germany, or Australia, you've got one national time series. Maybe two if you can split by region. Bayesian priors fill the gap — they let the model produce credible estimates even with limited data, because you're not asking the data to prove everything from scratch.
- Single-country analysis works — not just US multi-state
- Typical requirement: ~150 weeks of data
- Priors compensate for limited observations
- Still produces actionable confidence ranges
Every answer comes with 'how sure are we?'
Frequentist models give point estimates: 'TV drove 12% of sales.' Precise-sounding, but it hides how much the model actually knows. Bayesian models give distributions: 'TV drove 10–15% of sales, most likely around 12%.'
That's more useful. A channel with a wide range needs more data or a structured test. A channel with a tight range is ready for optimization. Hiding uncertainty doesn't make it go away — it just leads to overconfident decisions.
- Full probability distributions, not point estimates
- Wide range = test more before committing
- Narrow range = optimize with confidence
- Decision-makers see the real picture
Expert knowledge isn't thrown away — it's encoded into the model
You know things about your marketing that pure data can't capture. You know last year's TV campaign was poorly targeted. You know a competitor launched in Q3. You know branded search mostly captures existing demand.
Bayesian priors let you encode this knowledge. The model starts where your team is — not from a blank slate — and the data pulls it toward truth. This is why we call them 'expert models.'
- Current beliefs about each channel are encoded
- Model adjusts where data disagrees
- Setup: typically 2–3 sessions to define starting assumptions
- All assumptions are transparent and auditable
Every month of new data makes the model sharper
Each month, new data arrives and the model retrains. In Bayesian terms, last month's results become this month's starting point — the model doesn't start from scratch.
Over months, uncertainty shrinks, confidence grows, and recommendations get more precise. Channels you've been running for two years will have tight estimates. A channel you started testing last quarter will still be wider — and the model tells you exactly how much more data would tighten it.
- Monthly retraining with new data
- Previous outputs inform new starting point
- Uncertainty decreases over time
- Channels with more data get tighter estimates
If you're not in the States, you need a model built for your reality
Here's what nobody in the US talks about: most MMM tools assume you have lots of geographic variation. America has 50 states, 200+ DMAs, and a culture of geo-testing.
If you're running marketing in a single country — even a big one like Germany or the UK — you don't have that luxury. You've got one national time series. Maybe you can split by region, but it's not the same as having 50 semi-independent markets running different campaigns.
Bayesian modeling was designed for exactly this reality. It can produce robust estimates from a single market because it doesn't rely purely on data volume — it combines data with informed starting assumptions. That's not a compromise. It's actually a more honest way to build models. It just happens to also work when you don't have data coming out of your ears.
What 'priors' actually means — without the jargon
Forget the statistics terminology for a moment. A 'prior' is just a starting assumption. You're already making them — every budget decision is based on beliefs about what works.
When you allocated 40% to Google Ads, you assumed it drives significant value. When you kept TV flat, you assumed it's roughly performing as before.
Bayesian modeling just makes those assumptions explicit and testable. We sit down with your team, document what you believe about each channel, and then the model checks those beliefs against reality. Sometimes it confirms them. Sometimes it finds surprises. Either way, you end up with better answers than starting from a blank slate.
Frequentist regression
Works well with large datasets (50+ geographic units). Requires lots of data for stable estimates. Uncertainty is harder to quantify. Standard in the US.
Bayesian MMM (what Odins uses)
Works with single-market data. Combines expert knowledge with data. Full probability distributions. Built for the real world — not just the American one.
Platform attribution (Google, Meta)
Only sees its own channel. Uses its own attribution model. Systematically overclaims. Not independent measurement.