The GES Methodology

Macro, market and sentiment series are collected daily. Each is compared with its own previous 60 days: a signal counts as firing when it sits at least one standard deviation away from that baseline. For Supply, Behavioral and Monetary the direction that matters is learned from earlier events (after at least four), so a fall in an oil price no longer counts as an oil shock.

Each regime has a short, authored list of signals in three tiers (primary, secondary, contextual) weighted 1.0, 0.6 and 0.3. The score is the average of three measures: how much of the weighted list is firing, how persistently it has fired over the last 30 days, and how many different channels agree. The tiers and weights are judgments, not fitted to outcomes.

Forming means the score is at or above its own 80th percentile over the trailing 250 days; Severe means at or above its 95th. Thresholds lag by a day so a score never moves its own bar. A regime that is quiet for a year gets a lower bar than one that is often noisy.

A trigger is the first day the score rises through its Forming threshold after a cooldown equal to the regime's hold. The hold is fixed per regime, in calendar days:

Regimes overlap a great deal (on average 4.6 holds are active together), so OM Atlas keeps one live scenario at a time. Intensity is where a regime's score sits between its Forming and Severe thresholds. A new trigger replaces the live scenario only if its intensity is at least 0.10 higher; otherwise it is logged as context.

We measure each regime's forecast skill as the AUC of its score for events in its lookback window. It is always shown as a range with the date it was measured, because it is same-history testing and it moves: adding three months of data moved Supply from 0.81 to 0.67. Regimes below 0.60 are shown for monitoring only, with no forecast claim.

GES runs from 0 to 100 and is built from each company's annual filings. Four of eight planned dimensions (revenue, supply chain, manufacturing, customers) have filing-based data today; the average of the available dimensions is the base GES. Where a filing does not say, the value falls back to a competitor, sector or universe average, and the source is shown wherever a score is shown. Most values are still averages, which pulls scores toward the middle.

Inside a sector, lower-GES companies outperformed higher-GES ones by +84 basis points at 60 days and +133 at 90 days (2019 to 2026, t about 2), and that edge is modest, depends on the period, and is not explained by size, volatility or momentum. Across the whole market GES has no predictive power. That is why it is used only to rank companies inside a sector, and why it is never presented as a market signal.

For every earlier event of the regime, each sector's return over the hold is compared with the average of the ten sectors. The three worst sectors are marked Vulnerable for that event and the three best Beneficiary. Recent events count more, with a half-life of seven years. This gives, for each sector, how often it was Vulnerable and how often Beneficiary after this regime.

Some sectors are choppy all the time, so a raw rate misleads. Each sector is compared with its own unconditional rate; the ratio is its lift. A sector is used only with a lift of at least 1.3 and at least four prior events (High confidence from ten), up to three per side. A sector that appears on both lists is dropped.

This is deliberately demanding. Across 162 regime-to-sector tests only one link survives a correction for multiple testing, so most scenarios rest on the sector evidence plus the company ranking, not on a strong sector story.

Every holding falls into one of three groups: Tier 1V (its sector is on the trim list), Tier 1B (its sector is on the add list) or Tier 2 (everything else). Tier 1 names are scored with the GES re-weighted for the regime that triggered, using authored weights; Tier 2 names with the plain GES. Each name is ranked inside its own group (a group needs at least five names) and turned into a score from -1 (least exposed in the group) to +1 (most exposed).

In a trim sector, names above the group median are trim candidates, in proportion to that score; names below it are left alone. In an add sector, names below the median, the least exposed, are add candidates; names above it are left alone. Outside both lists only the extreme tails are touched (above the 90th or below the 10th percentile), rising from zero at that line, so there is no sudden jump.

Base trim = a fixed 10% knob × the company's conviction (0 to 1): at most 10% of a position is sold in one trigger. There is no client dial. The knob size was tested and only scales results linearly.

Momentum: if a stock the rule wants to trim has been rising strongly against its own sector (a blend of the 1, 3 and 6 month returns, weighted to the recent months), the trim is reduced, down to a floor of 0.35 of the original. A stock that is falling keeps its full trim. Size: a position above 5%, or one of the largest names that together make up the first 20% of the book, has its trim reduced in the same way, to the same floor.

The two are applied exclusively: the momentum factor where it reduces the trim, otherwise the size factor. Both can only reduce a trim, never enlarge it.

The freed capital goes to the add candidates (the add sectors and the low tail outside them), each in proportion to its conviction × its current weight, so every add rises by a similar relative amount. A position already above 5% is damped as it grows. What cannot be placed is held as cash, and the running example shows how little that is.

The position is held unchanged for the regime's hold. The same portfolio is then compared with and without the changes over that hold. Each event is its own scenario and is shown on its own.

• Thresholds and signal choices were developed on the same history they are measured on; skill has not yet been confirmed on live data. A prospective log starts at launch.
• Each regime has only a handful of independent events, so skill estimates move with new data.
• Most exposure values are averages; the monthly exposure table lags by a month or more.
• No transaction costs, taxes or liquidity limits are modelled; cash earns nothing.
• Past performance, whether actual or back-tested, is no indication or guarantee of future performance.