Custom Models
On this page
A custom model is a statistical model you specify in a conversation and Adverity Atlas fits on your data. Use it when the question needs a model shape the standard tools do not offer: how much volume a product loses when its price goes up, how a conversion rate differs between groups, or a media mix model where you want to choose the response curve or the priors yourself.
Like Marketing Mix Modeling, there is no separate screen. You ask in Chat, the model is fitted in the background, and the result comes back into the conversation.
Prerequisites
-
The Editor role or above to check or fit a model. A Viewer can read results, see Permissions
-
The data in a connected data source or an uploaded file. Adverity Atlas looks for a suitable table in the workspace first and asks you only when it finds none
Three ways to get a model
Scroll horizontally to see the full table.
Start from | Answers | Data it needs |
|---|---|---|
Price response template | How much volume each product loses per 1% price increase, and how much each promotion lifts sales | One row per product per period: units sold, net price paid, a 0/1 column per promotion. At least 30 rows with sales per product |
Media mix template | What each channel contributes, with the response curve, carryover, and priors you choose | One row per period: the outcome and one spend column per channel (up to 20). At least 26 periods, 104 recommended |
A model you describe | Anything you can state as a model: a regression pooled across groups, a count or conversion-rate model, your own carryover and saturation | Whatever the model refers to, one numeric, group, or date column per input |
Name the template, or describe what you want the model to capture, and Adverity Atlas proposes the model before anything is fitted.
Price response
Products are modelled together: a product with only a handful of price changes borrows strength from the rest of the category instead of returning a noisy number, and the result flags products whose price barely moved so you know which estimates are measured and which lean on the category. Distribution (the share of stores stocking the product), other controls, and media spend can be added as extra drivers, up to 40 in total, across up to 200 products.
Each promotion gets one lift shared by all products. Ask for a model you describe yourself if you need a lift per product. The model reports own-price elasticity at the prices seen in your data. It does not say where lost volume goes, to a sibling product or a competitor.
Media mix
Choose this over Marketing Mix Modeling when you want to set the form yourself: the shape of the response curve, fixed carryover or scale values for a channel, a log or a linear outcome, your own priors, seasonality, or a trend. Results of experiments or geo tests cannot be built into the fit, but Adverity Atlas can compare the fitted returns with them afterwards.
A model you describe
Describe the model in plain language. Adverity Atlas writes it down as named inputs, estimated values with their starting assumptions (priors), and one outcome, and shows it to you before checking it. A model can be hierarchical, for example a price sensitivity per product drawn from a category-wide typical value.
The model is built from a fixed vocabulary of distributions and functions, including saturation and carryover curves, and no code you or Adverity Atlas writes is ever run. Adverity Atlas refuses a description it cannot express, and says why. The limits:
Scroll horizontally to see the full table.
Limit | Value |
|---|---|
Estimated values and derived terms | 80 per model |
Data inputs | 60 per model |
Rows of data | 50,000. Aggregate in the query beyond that |
Distinct values in a group column | 2,000 |
Carryover window | 1 to 52 periods |
The data review before the fit
Adverity Atlas checks the model against the data before it fits anything and reports what it found, one question per check: is the data complete, are there unusual periods that could distort the fit, does the outcome move with each input, can the model tell the inputs apart. Charts of the data render under the review.
Scroll horizontally to see the full table.
Kind of finding | Examples | What happens |
|---|---|---|
Refused | Missing or non-numeric values, unreadable or repeated dates, a query result cut off at the row limit, a price at or below zero, a promotion column that is not 0/1, negative spend | The fit cannot start until the data is fixed. The review names the fix |
Needs your decision | Gaps in the periods, short history, little variation in price or spend, two inputs that move together, a spend column left out of the model, periods far outside the usual range | Adverity Atlas explains it and asks whether to go ahead |
Note | One to two years of history, promotion lifts shared across products | Passed on, nothing to decide |
The review also compares the table with the source it came from, and tells you about any periods that were dropped, values that were filled in or read from text, and columns a query created. Decide on those before you fit.
Fitting a model
-
Ask for the model, naming the template or describing what you want.
-
Adverity Atlas loads the data, runs the review, and asks Fit this model?, or Fit this model, accepting the warnings above? when there are warnings. Choose Change something to adjust the model or the data first.
-
The fit runs in the background, and a card under the request shows the model fitting. The conversation stays usable while it runs. It also shows under Tasks in the account menu, see Task Executions.
-
When it finishes, the card changes to a Show result button, a message states whether the fit converged, and you get a notification.
A fit is refused unless the same model on the same data was reviewed first, so settle the review before you confirm. Fitting stops at a 30 minute limit. A fit that hits it can be retried with a shorter sampling run, fewer values per group, or fewer rows. Re-running a failed fit is safe.
Reading the result
Show result renders the estimates and how well the model fits.
Every estimate is reported as a range rather than a single number, with how likely it is to be positive where that is meaningful. Where the data said little about a value, the result says so: that estimate mostly repeats the starting assumption rather than measuring anything.
The fit is given one of three verdicts:
Scroll horizontally to see the full table.
Verdict | Meaning |
|---|---|
Usable | The fit settled, and the estimates can be read as measured |
Directional | The fit mostly settled. Read the estimates as a direction, not a number |
Not converged | The fit did not settle. Do not use the estimates |
The charts that come with every result show the estimates, what drives the outcome, the return per unit of spend for spend inputs, and the observed outcome against the model's prediction. The templates add their own:
-
Price response: a demand curve per product, what a price change from -20% to +20% does to volume and to revenue, and the lift per promotion
-
Media mix: the response to a steady weekly spend per channel, and how long each channel's effect lasts
You can ask to reword the labels, descriptions, or units in the result, for example to use your team's names for the inputs. The estimates are not changed and nothing is refitted.
There is no score that compares two models. To choose between two specifications, fit both and ask for the results side by side.
Looking up an earlier fit
Ask about a past model in any conversation and Adverity Atlas retrieves it, including its status if it is still running. The notification for each finished fit links back to the conversation it ran in.
Permissions
Scroll horizontally to see the full table.
Action | Required role |
|---|---|
Read a custom model result | Viewer |
Check, fit, or relabel a custom model | Editor |
The Analyst role can chat about data but cannot check or fit a model. See Members and roles.
What's next
-
Marketing Mix Modeling, the standard media model, with budget optimization
-
Knowledge, record findings so the next model can build on them
-
Glossary, define the outcome measure before you model it
-
Artifacts, turn the result into something you can share
Adverity Atlas 2026.11.4. Powered by AI.