How the readings
are built

Crashboard turns published government data into two daily numbers. This is the essential logic: what the numbers claim, how unlike measurements are made comparable, and how the result becomes a verdict.

1What is being calculated

Two numbers, once a day, each running from −1 to +1.

Conditions Now is built from twenty indicators of what is happening at present, sorted into three groups: damage already done, markets under stress, and cash and funding.

Conditions Ahead is built from twenty-three different indicators of pressure building underneath, in five groups — exposure accumulated, the cost of money, borrowing accumulating, the nominal anchor, and uncertainty and unrest. It claims a horizon of six months to two years, and that claim constrains what may go in it: every indicator there is a state — a level measured against its own record, or a change over not less than a year. A four-week change cannot tell a world where nothing is building from one where everything has built and stopped growing, and the second is the condition this lens exists to see.

No indicator appears in both panels. Three series do — the ten-year Treasury yield, the broad dollar index and the OFR’s equity stress measure — and no panel reads the same thing from them. Conditions Now takes how violently a price moved and how closely it moved with everything else; Conditions Ahead takes where it has arrived. The equity series is a stress reading in the present panel and, in the forward one, only the interpolator that carries a quarterly market-capitalisation figure between its prints. When the two lenses disagree it is because different evidence points in different directions, not because one reading has been counted twice.

Each number is a weighted average of its own panel, and the average is then ranked against every day since 1998 to produce the word on the screen. Everything below is what happens inside that.

That ranking is what makes a historical reading something other than a record of what this instrument once said. The inputs are point-in-time: every indicator is normalised on an expanding window that ends on the day being scored, so no future observation reaches the raw number. The scales are not. A day is ranked, and its sigma taken, against the whole record — including the years after it — by today’s panel at today’s weights. A reading for 2008 is therefore where that day sits in twenty-eight years of hindsight, not what would have been printed in 2008, and the distinction matters most on the forward lens: because it has drifted lower over the record, earlier days rank calmer against the full history than they would have against their own. The present lens is near enough unaffected. Neither is a backtest, and neither is offered as one.

2The shape of it

Everything starts with published data, downloaded from public endpoints and stored with the units its publisher states. Nothing is bought, and nothing is scraped. From there each indicator is a formula over one or more series, producing one number a day — a difference between two rates, a ratio of fast movement to slow, a count of failures over a year. The formulas differ because the quantities do, and at this point nothing is comparable with anything else.

3Making unlike things comparable

This is the step the whole instrument rests on. Each raw value is expressed in standard deviations against its own history — everything observed up to and including that day, and nothing after it — and then bounded into −1 to +1.

Scoring against its own record is what lets a funding spread and a house-price change sit in the same average: each is measured against what is normal for it, so both arrive on one scale without anyone choosing a conversion factor. Bounding is what stops a single extreme reading dictating the answer. Here is one indicator all the way through.

4What a contribution is

A reading is not yet a contribution. Three things are applied to it.

Direction says which way is favourable, so both panels run in the same direction whatever the underlying series does. Weight says how much an indicator counts: three values exist — a half, one, and two — and which one an indicator carries follows from a written claim about what it measures. Three values rather than thirty-seven, because a weight for every indicator would be thirty-seven free parameters, and thirty-seven free parameters cannot be told apart from a model tuned until its output looked right. Confidence says what the reading is still worth given its age.

5Many numbers into one

The reading of the day is the sum of the contributions, divided by the weight that actually counted — each indicator’s weight times its confidence, so a stale indicator shrinks in the numerator and the denominator together and the average does not lurch when a slow series finally reports.

Alongside it the panel reports how much of itself agreed: the same weighted sum taken over the signs rather than the values, which is one when every indicator points the same way and zero when they are evenly split. That drives how saturated the colour is on screen, so a grey stretch means the indicators disagreed rather than that nothing happened.

6From a number to a verdict

A reading on this scale is not a communicable fact, so the headline is not the number but its rank against the record.

The cuts are in the record rather than on the scale, and that matters: the reading is free to run −1 to +1 but has never used most of that range, so thresholds placed on the scale would be unreachable at one end and constant at the other. The consequence to be clear about is that the ranking is always relative. There is no absolute zero and no calm decade — a quiet stretch still has a worst day, and it will be coloured like one.

7Whether the split works

Everything above describes how a reading is built. None of it establishes that the two lenses measure different things in a useful order — that is a claim about the instrument, and it can be tested. Conditions Ahead is supposed to register pressure accumulating; Conditions Now, what has arrived. If the split is real, then Conditions Ahead should predict future Conditions Now better than it predicts past Conditions Now.

That second half is the whole trick. Correlate the two series at any gap and some relationship appears, because both move slowly and any two readings a year or two apart resemble each other whichever came first. So the comparison is run twice: Conditions Ahead against a Conditions Now reading that came after it, and against one that came before. Predicting the past is impossible, so the second measures that resemblance on its own, and only what the first scores above it can be anticipation.

Read a row across and the first column of each pair wins, at every distance and in both halves of the record. At eighteen months Ahead predicts future Now at +0.503 against a control of −0.416 before 2012, and +0.623 against +0.159 after it. The margin is the finding. In the early half it widens all the way out, from 0.40 at six months to 0.98 at twenty-one where it settles, and the control there is not merely beaten but negative at every distance in the table — so an earlier reading of Conditions Ahead anticipates Conditions Now while a later one points the opposite way. In the later half the margin is widest near twelve months and narrows from there, to 0.23 at twenty-four. Every shift clears its control in both halves at once, which was not true of this panel before the autumn of 2026.

Read down and the two halves agree on the order but no longer on the strength, and the weaker of the two is now the early record. Squaring the eighteen-month figures gives about twenty-five per cent of Conditions Now’s movement accounted for in the first half of the record and about thirty-nine in the second — the later half is now the stronger of the two, having once been far the weaker. A year earlier the same comparison read under one per cent against twenty-one, and the early record carried almost no signal at all. Even so, the larger figure leaves well over half of Conditions Now unaccounted for, so this is a loose relationship reliably pointed the right way, and not a forecast.

Three limits. No month is privileged. The forward panel is built from quantities that move over quarters and years, so the broad plateau from eighteen months out is the shape to expect, not a horizon anyone picked. The agreement is between two measurements, not a measurement and an event. Conditions Now is itself a reading of conditions; tested directly against dated outside outcomes, neither lens beats its own base rate, and the section below says so again. And the sample is small where it counts. Three thousand paired days are not three thousand independent ones — the record holds a handful of cycles, and every figure above should be read against that denominator rather than the day count.

8What it does not do

It is not a forecast. Every step above is a statement about data already published. Ranking today against the record says where today sits; it says nothing about tomorrow.

It is fitted in exactly one place. No coefficient is estimated against a past crisis, no weight was chosen because it improved a backtest, and no cut is placed where it would have called a particular event. The exception is one indicator whose quarterly numerator arrives months after the quarter it describes; between prints it is estimated from daily public components by a regression refit at every release. It is corrected whenever the next print lands, it is discounted as the quarterly figure it stands in for rather than as a fresh one, and it estimates what that number is today — it says nothing about any day in the future. Against fifty-three real-time vintages it cuts the error against the next print from 8.83 per cent to 4.95, and beats holding the stale figure in 74 per cent of them.

It measures the United States. Both panels read one financial system.

It cannot see what is not published. Every series has to be freely redistributable, which rules out implied volatility and credit spread indices, and several indicators exist in the shape they do because of it.

The panel is not as independent as it is large. Indicators that look unrelated in calm markets converge under stress, so the agreement figure is more confident during a crisis than the underlying evidence strictly supports.

The early record is thinner. Fewer indicators reported in 1998 than today, and an indicator scored against three years of history is ranked more coarsely than one scored against thirty.

Data sources are attributed in the app’s about screen. This product uses FHFA data but is neither endorsed nor certified by FHFA. Further reading: the guide explains how to read the display; the today page publishes the current Conditions Now reading and the panel behind it.