Macro and plumbing scores

How the Macro Board and Monetary Plumbing build z-scores, section scores, factor cards, the reserves bridge and event studies, and what was tested.

The question

Is anything in funding, rates, credit or inflation unusual today, and do the other series agree? The Macro Board and Monetary Plumbing compute their scores from published series. This page describes how each number is built from those series, and what was tested.

The method

Features, factors and sections

Feature: one published series, or one spread or ratio of two series.
Factor: a group of features that measure one mechanism, such as reserve scarcity or the Treasury front slope.
Section: a group of the factors of one market.

Seven sections have scores: Funding, Fed corridor, SOFR trailing averages, Treasury curve, Real & inflation, Credit and Monetary plumbing. Two more Macro Board sections, Policy path and Releases, show data only. They have no factors and no score.

Z-score and percentile

A feature's z-score is its latest value minus the mean of its trailing window, divided by the window's population standard deviation. The window is 1,260 observations for a daily series and 260 for a weekly one. A window with no variation gives no z-score. The percentile ranks the latest value against the full history, with tied values counted as half.

A factor's z-score is the mean of its members' z-scores. Its directed score first multiplies each member by its stress direction, so a positive value always means more strain. A member with no stress direction is left out of the directed score. A factor with no directed member has no directed score.

Section scores

Stress is the mean of each directed factor's normal cumulative probability, times 100. A score of 50 is the five-year norm. Funding, Fed corridor, Credit and Monetary plumbing have a Stress score. A score of 84 means that, on average, the factors are about one standard deviation on the strained side.

Dislocation is the mean absolute z-score of the section's factors, ranked against the same mean on every past day of that section. The past values come from rolling z-scores, each using at least a year of observations. Real & inflation and Credit have no Dislocation score.

Coherence uses the sum of the positive directed scores and the sum of the negative ones across the section. It is 100 times their difference divided by their total, so 100 means every factor points the same way. It requires at least two directed factors with an absolute directed score of 1 or more. With fewer, the section shows N/A. The active factor count counts these factors only.

Model health scores the reliability of a section's statistics today. It combines data freshness, factor independence, half-life stability and regime fit. The calibration component has no measured value yet, so it is excluded and the score uses the other components. Health coverage is the share of the score's weight that had a value, so it is below 100 on every run.

Beside the scores, each section shows labels such as the curve regime and the corridor status. It also shows one sentence, chosen from a fixed table by the factors furthest from their norms. The measure pages list each label's thresholds.

Half-life and 21-day reversion

A first-order autoregressive fit over the trailing window gives each feature's persistence. The half-life is the log of one half divided by the log of that persistence. The 21-day reversion is the gap from the fitted long-run mean to the latest value, times one minus the persistence to the 21st power. Weekly series use 4 weekly steps instead of 21 days.

Factor cards

Monetary Plumbing shows nine factors on eight cards; the two corridor factors share one card. Each card's state comes from its factor z-score. That score is the directed score where one exists, else the plain mean. Bill cheapness and the bill curve have no direction.

Elevated: 1 or more; for an undirected factor, an absolute value of 1 or more.
Easing: -1 or less, directed factors only.
Weak: an absolute value above 0.75 and below 1.
Quiet: an absolute value of 0.75 or less.
Degraded: no member has a z-score this run. The card displays "insufficient data this run" and no score.

The funding anomaly maps each active factor to one mechanism. A directed factor is active when its directed score has an absolute value of 1 or more. Bill cheapness and the bill curve have no direction, so each is active when its plain z-score has an absolute value of 1 or more. When all active factors map to one mechanism, the label is that mechanism. Two or more mechanisms give Mixed or uncertain. A single mechanism becomes Calendar effect when two conditions hold:

An event falls within 2 market days of the Macro Board date.
At least two active member series have directed z-scores with the same sign as the event's mean effect, by majority.

Reserves bridge

The reserves bridge splits the change in reserves over 1, 4, 13 and 52 weeks. Change in reserves equals change in Fed assets, minus change in the Treasury's cash account, minus change in reverse repo take-up, minus a residual.

The window dates are weekly dates on which Fed assets, the Treasury's account and reserves all have a reading. Reverse repo take-up uses its last daily reading on or before each date. The residual is the part of the change that the three named inputs do not explain. It includes the liabilities that the bridge does not list, such as currency in circulation.

Event studies

The event studies cover six kinds of event: month ends, quarter ends, tax dates, debt limit binding, technical adjustments and facility starts. Month ends, quarter ends and tax dates come from the calendar. The other kinds are detected from the data:

Debt limit binding: 10 consecutive Treasury records with headroom under 0.20% of the limit.
Technical adjustment: the rate on reserves or the reverse repo rate changes, and the target range is unchanged for 2 days either side.
Facility start: the first date the rate on reserves or the reverse repo rate appears after the history begins.

The study covers each event and each daily plumbing series. It takes the change from the event day to each day from 10 before to 10 after. The average path plots the mean, the median and the 25th to 75th percentile band. The mean effect and the hit rate use the change 5 days after the event. An event kind needs at least 8 past events. A series needs 8 events with readings at day 0 and day 5.

Releases

The release rows show the current FRED value of each monthly release, its level and its monthly and yearly changes. A revision replaces the value in place.

Inputs and lag

The daily run ingests the FRED series and the Treasury debt data after every symbol, then computes the scores. The daily dates follow SOFR's observation dates. An observation of another series on a date with no SOFR observation is dropped. Weekly balance sheet series have Wednesday dates and change once a week.

Each value shows the date of the observation it uses. FRED publishes a day late, so the latest Macro Board date is usually one trading day older than the options data. A section whose date is more than 5 calendar days before the Macro Board date is marked stale. The lag rule in the daily run and the lag rule applies to model features. These scores describe the published data as of its own date.

How it was validated

21-day reversion

The 21-day reversion is tested walk-forward on each feature's own history. At each test origin the model is fitted on only the trailing 504 observations of a daily series (104 for a weekly one). It forecasts 21 days ahead, then moves 21 days on, so the forecasts never overlap. The model must beat four baselines on mean absolute error:

no change;
full reversion to the window mean;
the window's average 21-day change;
a z-scaled forecast.

Over all origins the model must also be right on direction more than half the time, with at least 8 tests. The origins then split at the median of their recent volatility. Each half with at least 8 tests must beat the same four baselines. A feature that fails these gates shows no 21-day reversion value. A feature that passes in one half only shows its value with a regime note. The interval shown is an 80% Student-t interval on the walk-forward errors.

Half-life

The half-life must first pass its own checks. The fit needs at least 504 daily or 104 weekly observations, and a persistence below 1 at 95% confidence. Each half of the window must refit to a half-life between half and twice the full value. The half-life must also be under a quarter of the window.

Audit of 29 July 2026

The audit ran these gates on the 76 Macro Board features. 19 showed a half-life and none showed a 21-day reversion. Of the 17 features that reached the baseline test, 16 lost to the no-change forecast, 12 by more than 10% in mean absolute error. The audit found the gates sound: no tested relaxation showed a value on more than 4 of 76 features.

Output contracts

Each run produces three outputs: the feature statistics, the section scores and the event paths. The caller checks all three against written contracts before it installs any of them. When a check fails, the previous files stay in place.

Untested outputs

The section scores, the factor states and the event studies have no validation test. They compare each series with its own history and summarise past events.

Limits

The scores do not suggest a position and are not probabilities of future events.

The z-score has no minimum history. A series with a short history is scored against fewer observations than the window length, and so is a factor built on it.

The walk-forward test fits 504-observation windows. The 21-day reversion shown comes from the fit over the full 1,260-observation window.

A section scored on few factors moves further per factor than a section with all its factors. A degraded plumbing factor indicates missing data.

The reserves bridge residual includes every gap between the four inputs, including differences in how each input is dated within the week.

The release rows show the current vintage only.

Sections follow different release calendars, so two sections on one screen can describe different days. The section date and the stale mark show the day each section describes.