Heatmap: an Offside column, so the grid says who is losing

[mspinola]

What this adds

An Offside column on the heatmap, in its own "Cost Basis · Large Specs" group: how far Large Specs sit from their own average cost, in the market's weekly standard deviations. Negative is under water. Reads cotmetrics.offside (mspinola/cotmetrics#28); this repo computes nothing.

Every positioning column on this page is about size — how unusual, how extreme, how many dollars of risk. None said whether the holders are up or down.

Four decisions worth reviewing

Its own group, not a third Exposure column. It reads a different cohort (LEG_LARGE, where the dollar-risk column uses LEG_SPEC) and answers the opposite question. A basis computed on the summed large+small net describes a trader who is both cohorts at once, and the two have different average costs and behave differently when under water. Large Specs is also the cohort every published figure for this measure is quoted on.

No percentile, unlike Risk %ile. Dollar risk is incomparable across markets, so it must be ranked against a market's own history before it means anything. Offside is already comparable by construction: dividing by the market's own sigma is what the measure does, and 0 means "at the cohort's average cost" in every market. Ranking it would replace a readable quantity with a percentile of one.

Only the losing tail lights, and the colour is a P&L statement. Lit at <= -2 sigma in the bear colour. That is the sign of the cohort's own mark, not a direction verdict: red means "these holders are down", not "this market goes lower". Being deep in profit is not distress, so the column is deliberately asymmetric where a z-score column would not be. -2 is a display threshold, deliberately rounder than any figure in the study (pooled p10 is about -1.7).

The tooltip refuses to promise capitulation. The obvious inference from a lit cell — they are trapped, so they must fold — was pre-registered, tested against 44 markets and did not hold (npf#214, verdict "adverse-move proxy"; within equally severe adverse moves, deep-offside cohorts capitulated no more often and Small Traders significantly less). The copy says the number is a reading of who is losing and not a forecast, and two tests pin that wording so it cannot quietly drift into a claim.

Verification

• Live store, all 18 rows in Energies/Metals/Currencies populate with consistent basis/mark pairs (Euro reads -1.7 because Large Specs are net short into a rising euro; Mexican Peso +5.9 long and in profit).
• The lit path, end to end: driving the real render_heatmap_layout at target date 2026-06-30 returns Euro -2.09 and Corn -5.91, both matching the emitted condition params.value != null && params.value <= -2.0 and rendering in #FF4D4D.
• In the browser: group header, +/- formatting and dim styling all render as intended; screenshot behaviour matches Risk %ile.
• Page registry still loads all 16 routes, which is the check that matters when use_pages imports every page at startup.
• 11 new store-free tests (join on own week, unpriceable market stays a row, failure returns None, object dtype for the null guard, styling both branches, and the two copy pins). Full suite 448 passed, ruff clean, check_dep_floors.py OK.

Floor

cotmetrics[options]>=0.7.0. This is an import-time floor rather than a nominal one: against 0.6.0 the module does not exist and, because use_pages imports every page at startup, the ImportError takes down the whole route registry rather than one column.

Merge mspinola/cotmetrics#28 first, or CI here cannot resolve the floor.

🤖 Generated with Claude Code