Built for thin data.

Marketing data is one point per week and ten channels to estimate. A blank-slate regression overfits. Odins starts from what you already know and lets 2 to 3 years of evidence correct it.

Book a demo How we set priors

Channel effect estimate

prior → posterior

The model starts from your team's knowledge, then lets the data correct it.

The data problem

Why blank-slate models struggle here.

Geo-split models work well when a market divides into dozens of regions. A single-country time series offers no such variation, so the modeling has to work harder.

One data point per week

Two to three years of history is roughly 150 weekly observations. That has to carry ten or more channels.

Channels move together

Campaigns launch with seasons and budgets rise in tandem. Separating effects needs more variation than one market offers.

Overfitting looks like insight

With thin data, a flexible model happily explains noise. The numbers look precise and mean nothing.

Prior-setting as a discipline

We don't start from zero.

Before the model sees a single week of data, it already knows what your team knows. Three sources:

Historical spend: reasonable levels and marginal return per channel
Saturation signals fitted from your digital channels' conversion and impression data
Structured interviews: "if you raise this channel 25%, what happens?" becomes a prior on where it saturates
Then 2 to 3 years of history (roughly 150 weeks) act as experiments that correct the starting point

Under the hood

For the readers who want the mechanics.

The full specification is shared with your team during onboarding. The short version:

Hill saturation functions per channel, time-varying so curves can shift with seasons and pricing
Channel-specific adstock: carryover and lag estimated per activity type, not assumed
Control variables for seasonality, price and promotions, so channel effects aren't polluted by them
Full Bayesian inference: thousands of sampling iterations per training run, posterior uncertainty on every parameter
Out-of-time verification: forecasts are made from a cutoff and scored against the actuals that follow

Channel contribution, 80% ranges

from the posterior

Search 14 to 16% optimize

Social 9 to 13% optimize

Display 2 to 9% test first

A tight range is a decision. A wide range is a test plan, with cost boundaries.

Go deeper.

The technical layer under every number in the platform.

The model

What gets estimated: baseline, saturation, adstock and controls.

See the methodology

Bayesian approach

Why ranges and monthly updating beat point estimates.

See the approach

Staying accurate

Daily data, monthly retraining, verification against actuals.

See verification

Data integration

600+ sources, offline included, structured into one dataset.

See data integration

One question to ask any vendor

How do they set their priors, and how does the model behave when data is thin? The answers separate MMM built for your market from MMM built for a bigger one.

FAQ

Frequently asked questions.

Do we need engineering resources?

No. Odins handles integration, modeling and maintenance. The typical customer effort is 2 to 3 one-hour sessions during setup.

What if our data is messy or incomplete?

Expected. The platform is built for real-world gaps: offline channels on templates, history loaded as-is, quality monitored continuously.

What exactly is a prior?

A documented starting assumption, for example where a channel saturates. You see and approve every prior before the model runs.

What if our starting assumptions are wrong?

The data has the final say. The weekly history acts as experiments: where the evidence disagrees with a prior, the model moves.

Do the priors stop mattering over time?

Largely, yes. After about a year of monthly retraining the model is driven mostly by your data.

How is the model estimated?

Full Bayesian sampling rather than a single fitted line: thousands of iterations per training run, so the uncertainty you see is the model's actual belief, not an afterthought.

Does this work for smaller markets?

It is built for them. Single-country time series are exactly where disciplined prior-setting earns its keep.

Who owns the data and the model?

You do. Structured data can be exported to your own warehouse at any time.

Do the saturation curves stay fixed over the training window?

No. Each channel's curve moves with the season, so the same spend can earn more in peak season than in the low season. How it works, what it costs and where the limits are: time-varying saturation.

How are brand and long-term effects modeled?

As an optional component. Each channel sends a share of its effect into a shared stock of brand value, which pays out sales over months and years and leaks unless it is refilled. The equations, the priors and the limits: brand in marketing mix modeling.