EDA (exploratory data analysis) is a quick sanity check on your data before you commit a training run to it. It has nothing to configure — connecting it makes it ready, and it runs when you press Run like everything else on the canvas. The difference is the wait: it finishes in the time the page takes to respond, not as a background job.
Connecting it
Drag an EDA node off Period (or directly off Variables to use the full date range). There's no form to fill in — the connection itself is the configuration. Clicking an unconnected EDA node opens a short explainer of what it does, with a Got it button, not a form.
What it shows
The results are one page of cards, not tabs. The analysis covers your Output, Media Spend and Organic columns; Contextual, Controls, Geo and Population don't enter any of the charts.
| Panel | What it's for |
|---|---|
| VIF (Variance Inflation Factor) | Flags channels that move together closely enough to confuse a model about which one is actually driving the outcome. A high VIF on two channels is a warning that the model may struggle to separate their individual contributions. |
| Correlation matrix | How every pair of those columns correlates with each other and with your output. |
| Time series | Your output and each channel plotted over the period, so you can spot an obvious data problem (a gap, a flat line, a spike) before training. |
| Spending Variability vs Share | Each channel by its share of total budget (x) and how much its spend swings from period to period (y), with bubble area showing total spend — useful for noticing a channel spent so flat there's little for a model to learn from. |
| Lag Analysis | Whether a channel's effect on the outcome shows up with a delay, which is a hint about how much adstock (carryover) that channel might need. |
In a by-market project the page adds a market selector and shows one market at a time — or the country-level aggregate.
Why run this before a model
Nothing here stops you from training anyway, but a few minutes here can save a much longer wait: two heavily collinear channels, a channel with almost no spend variation, or a data gap are all things that make a model's results hard to trust even when training finishes without error. EDA exists to catch those before you spend a training run on them.
What connects here
Upstream: Period or Variables. EDA doesn't have an output — it's a dead end on the canvas, not a step other nodes build on.