Methodology
How Jodie works — and how we know it works
Markets move in blocks. Stocks that share a story trade together — often across sector lines, where sector-based tools can't see the connection. Jodie measures that group behaviour directly: it subtracts what the market did, maps which names are moving together beyond it, flags newly forming groups with statistical evidence, and tracks each theme through its life. This page explains the machinery, the safeguards that stop it fooling itself, and the testing we did before asking anyone to trust it.
How the engine works
Four layers, each building on the last. Everything downstream inherits the first one.
Subtract the market first
Most of what any stock does on a given day is just the market. Before Jodie measures anything — a move, a group, participation — it removes the market's contribution. What remains is the beyond-market move: the part specific to the stock or the theme. Every number on the radar is built on this, which is why a broad rally doesn't light the board up with false themes.
For the technical reader Returns are residualized against an equal-weight universe proxy with shrunk rolling betas; intraday and daily layers use the same convention so live and historical analysis are directly comparable.
Map the standing structure
Each night Jodie rebuilds a map of which names durably travel together across ~1,900 US stocks — the long-run relationships that define the market's standing blocks. This baseline is what makes “new” detectable: you can only spot a relationship forming if you know what the old relationships were.
For the technical reader Correlation structure is estimated with Ledoit–Wolf shrinkage, denoised via random-matrix (Marchenko–Pastur) filtering, and clustered with Louvain community detection. Cluster identity is tracked day-to-day so themes have continuous lineages.
Detect what's newly forming
The interesting moment is when a group of names starts trading as one block that didn't before — often spanning several sectors, which is exactly why sector scanners can't see it. Jodie compares each pair's recent beyond-market co-movement against that pair's own history and flags groups only when the evidence is statistically real, not a one-day coincidence.
For the technical reader Short-window EWMA residual correlations are tested against each pair's own baseline via Fisher z, sized to the effective sample. Groups must form dense subgraphs with a high-evidence seed clique; acceptance thresholds are calibrated on null simulations to roughly one false formation per two trading weeks across ~1.8M candidate pairs.
Track each theme's life
During the session the engine re-reads the market every 15 minutes: how many members of each theme are participating beyond the market, whether that breadth is widening or contracting, and where the theme sits in its life — setting up, activating, expanding, or weakening. Themes are born, broaden, and fade; the radar shows you where in that story each one is.
For the technical reader Participation breadth is computed on residual returns (0.5 ≈ market-neutral), with volume, flow, and dispersion tracked per cluster at 15-minute cadence against 30-day per-name baselines.
Built to not fool you (or us)
Correlation tools fail in known, predictable ways. Each of these failure modes is engineered out of the radar — not as a disclaimer, but as machinery.
On a +2% market day, everything co-moves. Because every measurement is beyond-market, broad tape moves don't register as theme activity — only group behaviour the market doesn't explain does.
Earnings weeks make unrelated stocks crash together; index events synchronize names that have nothing in common. Every flagged pair must keep its co-movement after its single biggest shared day is removed, and extreme days are capped before anything is measured. A theme has to show up again and again, across many days, to exist on the radar.
With nearly two million stock pairs, coincidences are guaranteed. Detection thresholds were set on simulated markets containing no themes at all, and tuned until the machinery produced almost none — so when the radar shows a formation, it cleared a bar that random data essentially never clears.
Theme names are generated from what the member companies actually do, and every group ships with its evidence attached: how strongly it co-moves now, what its baseline was, and how many sectors it spans. When a group is a muddle, the radar says so rather than dressing it up.
The theme lifecycle
Every active theme is placed on a lifecycle that describes how broadly its members are participating beyond the market — from a narrow early move, through wide participation, to dissolution. This describes the present state of the group, not what comes next.
A group has just started moving together beyond the market. Participation is narrow; the formation could broaden or dissolve.
More members are joining the move. The group is no longer one or two names — confirmation is appearing across the block.
Participation is wide. Most of the theme's members are moving together beyond the market — the theme is fully visible.
Participation is contracting. Members are dropping out of the shared move and the block is losing its signature.
How we validated it
Most market tools ask you to take their word. We ran the test most vendors won't, and we publish the answer either way.
Before launch we replayed the full engine over four separate market regimes — a stress window, a choppy range, a calm low-volatility stretch, and a mixed quarter — measuring every output against what the market itself did over the same hours, with the luckiest trades removed, on data the engine had never seen.
The radar finds real structure, early. It flagged the AI-datacenter complex trading as one block weeks before that had a name, and tracked the defensive-rotation block through an entire selloff — both verified to be genuine multi-week co-movement, not one-day artifacts.
Used as an automatic trading signal, the engine's outputs carried no reliable forward edge once market exposure and luck were stripped out. So we don't sell predictions — and we publish that finding rather than hide it. Any service claiming otherwise should be asked for the same test.
Every claim Jodie makes on screen is one that survived this process: co-movement that is measured, evidence that is statistically sized, timestamps that are public. We sell sight — early, measured, and honest about its limits.
For the technical reader Validation protocol: walk-forward replay at 8 frames/day over four regime windows (one held fully out-of-sample), all outcomes measured market-adjusted (SPY-hedged, net of modelled costs), trimmed of the top 1% of winners to remove tail luck, with pre-registered hypotheses. Detection robustness: planted-theme recovery and null-market false-positive calibration; every published formation survives a drop-largest-shared-day test.
What the radar is — and isn't
Reading the radar well means knowing what it claims and what it deliberately does not. These are not legal caveats — they shape how the product should be used.
Jodie is a descriptive radar, not a forecaster. It shows which themes are active right now, which names are central to each, and whether participation is widening or contracting. It does not estimate price targets, the size of a move, or whether a theme continues.
The “leader” of a theme is the most structurally central name, not a recommendation. The names listed alongside it are the others currently moving with the block — exposure context, not a watchlist of what happens next.
Evidence labels (emerging / strong / exceptional) describe how statistically unusual a group's co-movement is versus its own history. They are never probabilities of future returns.
Theme detection is a lens on market structure. Use it to see what is forming, what you are actually exposed to, and how participation is changing — then bring your own judgement about what it means.
See it on today's tape
The fastest way to judge the methodology is to watch it describe the market you're already trading.