-1.png?width=1132&height=292&name=Asset%201@2x%20(12)-1.png "Odins.ai Logo")
By the Odins team · Updated October 2026 · About 6 minutes
Every year ends with the same question, asked in hindsight: with the budget we had, what should we have done differently, and what would it have given us? Most teams answer it with opinions. A marketing mix model answers it with a scenario: the same total budget, the same period, allocated the way the model would have recommended, forecast with a range. This guide explains how that works, how to read it, and how to turn it into next year's plan without turning it into a blame exercise.
The short answer
Take the trained model, give it last year's total budget and period, and let it allocate the spend across channels and months by marginal return. Compare that allocation with what was actually spent, and compare the forecast revenue of each. The difference, with its confidence range, is what the year could have given. Typically it points at the same two things: money left in channels that had saturated, and channels with headroom that never got it.
What the comparison actually shows
The model has a response curve for every channel: how much incremental revenue each level of spend returns, with diminishing returns as the channel saturates. Last year's actual spend sits somewhere on each curve. The optimised allocation moves spend from where the slope was flat to where it was steep, until the next unit earns the same everywhere. Three things come out of it.
- Where you were over-invested. Channels where actual spend sat past the point of diminishing returns. The last part of that budget returned less than it would have elsewhere.
- Where you were under-invested. Channels where the curve was still steep at the spend level you ran. More there would have returned more than the average.
- The timing. The same comparison by month shows where spend was flat across the year when the curves say it should have followed the season, with more in the months when each unit worked hardest.
| Column | What it contains | How to read it |
|---|---|---|
| Actual | Last year's spend per channel and month, and the model's estimate of the revenue it drove. | The baseline for the comparison. The model's view of last year, not the platforms' view. |
| Recommended | The same total, allocated by marginal return, and its forecast revenue. | What the model would have proposed, given what it knows now. |
| Difference | Revenue gained or lost, with a range, and the budget moved per channel. | The size of the opportunity, and where it sits. Read the range before the midpoint. |
What the comparison judges is habits
Nobody allocated last year's budget with this year's model. The comparison is useful because it shows which habits cost money: a channel kept at its contracted level, a flat monthly split, a reflex to fund what reported well in its own dashboard. Those are easy to change for next year. Treat the exercise as a plan, and the room stays on your side.
Read the range, then the midpoint
The gain the comparison reports is a forecast of a year that did not happen, so it carries uncertainty from every part of the model. Two cases deserve care.
- Channels that never varied. If a channel ran at the same level all year, the model knows its curve only near that level. A recommendation to double it rests on the prior and on what similar channels do, and the range will say so. Treat it as a test to run, not a move to make.
- Channels that were never tried. The model cannot credit a channel it has not seen. If the comparison shows no gain from, say, radio, it is because there was no radio in the data, not because radio does not work.
Across customers, moving spend from saturated channels to channels with headroom typically gives 5 to 15% more effect from the same budget. If your comparison shows far more than that, look at the ranges and at which channels carry the gain before you present it.
Turn hindsight into next year's plan
- Start next year from the recommended allocation. Next year's budget discussion usually begins with last year plus a percentage. Begin with last year reallocated instead; it is the same money, better placed.
- Keep the constraints that are real. If TV is contracted or a sponsorship is fixed, lock it in the scenario. The model then optimises the rest and shows what the lock costs, which is useful when the contract comes up for renewal.
- Schedule the tests. Every channel where the range is wide becomes a structured test in the plan: which channel, how long, at what cost, and how much narrower the range will be afterwards.
- Store this year's forecast now. Then next year's version of this exercise compares actuals with a forecast made in advance, which is a far stronger basis than hindsight alone.
Where Odins fits
In Odins the comparison is a scenario on the trained model: last year's total budget and period, allocated by marginal return across channels and months, next to the actual spend, with the forecast difference and its confidence range. Locked channels and ranges per channel are set in the scenario. The model behind it is retrained monthly by our team, every forecast is stored before actuals arrive and verified in the platform, and scenarios are self-serve, in the platform or from your own AI assistant through the Odins MCP. The same scenario, run forward, is next year's plan.
Frequently asked questions
Can a marketing mix model tell us what we should have done differently last year?
Yes. Run last year's total budget through the model's optimiser for the same period and compare the recommended allocation with the actual one. The forecast difference, with its range, is what the year could have given. The result points at saturated channels that were overfunded and channels with headroom that were not.
How reliable is the number?
As reliable as the model is for the channels involved. Where spend varied over the year, the curves are well placed and the range is narrow. Where it did not, the range widens and the recommendation is better treated as a test. Read the range first; a midpoint without one is not a result.
Does this blame the team that set last year's budget?
It should not, and it does not have to. The team did not have the model. The comparison shows which practices cost money, such as a flat monthly split or budget following platform-reported ROAS, and those are what change for next year.
Can we see what a different total would have done, not just a different split?
Yes. The same scenario with a larger or smaller total shows the forecast revenue at each level, so you can see whether last year's budget was the right size as well as whether it was in the right places.
The forward-looking version of the same question, how much it takes to reach a target with a given certainty, is in how much marketing does it take to be 80% sure of hitting a revenue target; the mechanics of allocation are in allocating budget across channels. How to take the result into next year's budget conversation is in building a business case for a bigger marketing budget.