The Deep Dive panel

Scroll past the Strategy Catalog and you reach Deep Dive · Cohort Shape. It answers a different question than the table above it. The table shows you individual strategies. The Deep Dive describes the cohort — the top N strategies that match your current filters, as a group.

That distinction is the whole point. One strategy at the top of a sorted list tells you almost nothing (something always wins by luck). The shape of the group around it tells you whether there's a real pattern underneath.

The Cohort selector

The panel's header states exactly what it's describing: "Every chart describes the top N strategies by MAR that match your filters" — and it says plainly that with a catalog this size, the top of any performance sort partly reflects selection bias. The Cohort dropdown sets N: 300 / 500 / 1,000 / 2,000.

Use it as a stability check: if the picture changes completely between a 300-cohort and a 2,000-cohort, what you're looking at is thin. If the same structure holds as the cohort widens, it's a property of the strategy family rather than of a handful of lucky variants.

The charts also follow your filters. Filter to MEIC only and you get the MEIC cohort's shape; add a stop-loss chip and the panel re-describes that slice.

In Advanced view

The panel follows the Simple/Advanced toggle like the rest of the Browser. Switch to Advanced and every chart title, the header sentence, the Cohort tooltip, the Quality Curve's axes and both heatmap legends change their metric word from MAR to wMAR — Width × Stop-Loss · Avg MAR becomes Avg wMAR, and so on.

That is not just relabelling. The cohort itself is fetched at a different weight: in Simple the panel ranks at base weight 0, which collapses the blend to the reference period's MAR alone, so the values genuinely are MAR. In Advanced it ranks at your actual Base Weight, so they are true wMAR. The label and the ranking move together by design — a panel that renamed the metric without changing the ranking would be lying about what it shows.

Practically: the cohort you see in Advanced can contain different strategies than the one you saw in Simple, because it was ranked by a different number. If you switch views and the picture changes, that is the weighting talking, not the data. wMAR & reference periods explains the blend itself.

One asymmetry worth knowing: the Robustness scatter plots raw MAR Base against MAR Reference on its axes in both views, because those axes are the comparison — the blended metric would defeat the chart's purpose. So in Simple view you will see a chart naming a base period that the Simple period selector doesn't show you. That is the chart being literal, not a bug.

The four charts

Entry-Time Distribution

A bar per entry time, counting how many of the cohort's strategies enter then.

Reading it: peaks are the times of day that dominate the top of your current sort. A cohort clustered into two or three narrow spikes means the edge in your filter set is concentrated in a few minutes of the day — worth knowing before you build a portfolio that unintentionally stacks every position into one moment (and see Min entry spacing for the correlation cost of that). A broad, even distribution means entry time isn't the deciding variable in this slice.

Robustness · Base vs Ref

A scatter plot: each dot is a strategy, positioned by its performance in one window versus another, with the dot size showing trade count and a diagonal line marking "both windows agree".

Reading it: dots near the diagonal performed consistently across both windows. Dots far above it did well recently but not over the longer window — a recency story, not a durable one. Big dots are backed by more trades and deserve more weight than small ones. What you want for a portfolio building block: on or near the line, and large.

Quality Curve

The cohort's ranking metric sorted from best to worst, drawn as a curve — how concentrated the edge is.

Reading it: a steep cliff at the left means a handful of variants tower over everything else — the edge is concentrated, fragile, and probably part luck; you'd be betting on those specific parameters. A flat, gentle slope means dozens of comparable setups performed similarly — a robust plateau, where picking a strategy from the middle costs you little and buys you a lot of safety. Prefer plateaus to cliffs.

Parameter heatmaps: Width × Stop-Loss and Width × Premium

Two grids. Each cell is a parameter pair; the shading is the cohort's average performance there; the big number is that average and the small number is how many strategies sit in the cell.

Reading them: look for contiguous warm regions, not the single hottest cell. A block of neighbouring cells that all perform decently is a parameter region that worked — land in the middle of it and small errors don't hurt you. An isolated hot cell surrounded by cold ones is the visual signature of overfitting: the exact same thing Confluence measures per strategy, seen from above. Also watch the small numbers: a spectacular average over three strategies is noise.

A worked example

Say you filter to MEIC, 100-wide, stop 150%, and sort by MAR. You could take the top row and move on. Instead, glance down:

1. Quality curve — steep cliff? The top row is a lottery ticket; widen the cohort to 1,000 and see whether the cliff survives.
2. Width × Premium heatmap — is the top row's cell part of a warm block, or an island? An island means its neighbours (a nudge in premium either way) didn't work, so the result rests on a coincidence.
3. Robustness scatter — is it on the diagonal, or a recency spike?
4. Entry-time distribution — is your shortlist all one time of day? If so, you've built one bet, not five.

Then go back to the table and add the strategies that sit inside the healthy regions. The Deep Dive doesn't tell you what to trade — it tells you which parts of the sorted list are worth trusting.

An honest limit

These charts describe the cohort's historical shape, over the period you selected, for the filters you set. They are not a recommendation, and a handsome cohort shape is not a promise: it's the absence of one specific failure mode (isolated overfit), not the presence of a future edge.