Market data · Recession probability since 1967
Recession probability
The recession probability model reads 0.76% for July 2026. That is an estimate of whether the economy was already in a recession during that month, which is a different question from whether one is coming, and the difference is most of what this page is about. Across 710 months since June 1967 the model has been right about 62 of its 64 alarms and it warned about 0 of 8 recessions a year ahead. Both numbers are computed below from its own record against NBER.
Model output for July 2026, published 2026-09-09 710 monthly readings since June 1967
Probability the US was in recession
0.76%July 2026. The 75 months in a row under 10%
Where that sits on 59 years
Read this before the chart
This model estimates whether a recession is happening now, from activity that has already been measured. It is not a forecast, and the record of every recession below shows it reading under 10% twelve months before all 8 of them.
The numbers
710 monthly readings, checked against NBER
Every figure here is scored against NBER's business cycle dating, which is the only authority on when a US recession happened. The probability comes from July 2026, the last month the model covers, because the four activity series behind it publish on a lag of roughly two months.
Recession probability
0.76%0.58% in Jul 2025
The model's estimate for July 2026 itself, not for a month ahead of it. A change here is measured in percentage points rather than percent, so the card prints last year's level beside it instead of a return.
Months under 10% that were recessions
0 of 613NBER dated 0 of them a recession month
NBER has never dated a month as part of a recession while this model read under 10% for it. That is 613 months of agreement and it is the strongest thing on this page.
Months above 40% that were recessions
62 of 64NBER dated 62 of them a recession month
The other end of the same test. The 2 exceptions all sit within 3 months of a recession NBER did date, so the model was early rather than wrong.
Recessions it warned about a year ahead
0 of 8highest reading a year out was 1.62%
Twelve months before each of the 8 recessions in the record the model read under 10%, every time. It describes the month it is dated to and nothing beyond it.
Months since the last reading above 10%
75last was April 2020
The current quiet run. 86.3% of the whole record sits under 10%, so a long stretch down here is the normal state of this series rather than a signal in itself.
Record since Jun 1967
710 months85 of them dated as recession
The high is 100.0% in March 2020 and the low is 0.04% in September 1983. The median month reads 0.44%, which is why the average of 8.55% describes almost none of them.
Sources: Chauvet and Piger smoothed US recession probabilities, from four monthly coincident activity series. Federal Reserve Bank of St. Louis, NBER based recession indicator for the United States. LBMA Gold Price (PM London auction, USD/oz). Retrieved 2026-09-09
The long record
Recession probability chart since 1967
The shaded bands are the 8 recessions NBER has dated inside this record. The line spends 86.3% of its life under 10% and almost all of the rest inside a band, which is what a well-behaved coincident indicator looks like. What the chart cannot show is the part that matters most, which is that the line is still flat on the left edge of every band.
Smoothed probability the US economy was in recession, monthly
Recession probability, 1967 to today
July 2026
0.76%Median month reads 0.44%
The dashed rule is the 8.55% average across all 710 months, which describes almost none of them, because a series that sits near zero and then jumps to 100% has no typical value. Turn on the shading to see the 8 NBER recessions drawn behind the line.
Source: Chauvet and Piger smoothed US recession probabilities, from four monthly coincident activity series. Recession dating from Federal Reserve Bank of St. Louis, NBER based recession indicator for the United States. Retrieved 2026-09-09
Download the full dataset as CSVWhere the 710 monthly readings landed
Distribution of the recession probability
613 of the 710 months read under 10%, so this series is a flat line interrupted by spikes rather than something that drifts. The middle bands are nearly empty at 33 months between them, because once the model starts moving it moves fast. Today's 0.76% sits in the dark bar.
What it means
0 of 613 months under 10% were recessions
Almost always, about the month it is describing, and almost never about the month after. Sorting all 710 readings into bands and asking what NBER later said about each one produces two different answers depending on which question is asked, and the gap between them is the whole subject of this page.
| Reading | Months | Episodes | Was a recession month | Recession within 12 months | Median months to it |
|---|---|---|---|---|---|
| Under 10% | 613 | 12 | 0 of 613 (0%) | 88 of 602 (15%) | 7 |
| 10% to 20% | 11 | 9 | 4 of 11 (36%) | 6 of 11 (55%) | 1 |
| 20% to 30% | 11 | 6 | 9 of 11 (82%) | 8 of 11 (73%) | 1 |
| 30% to 40% | 11 | 7 | 10 of 11 (91%) | 10 of 11 (91%) | 1 |
| 40% and above | 64 | 10 | 62 of 64 (97%) | 61 of 64 (95%) | 1 |
The fourth column is the model's own claim. The fifth answers a different question and is computed only over the 699 months whose following year NBER has finished dating, which is why it is the smaller count
The same table, drawn
Hit rate by reading band, both questions
The dark bars climb from 0% to 97%, which is what calibration is supposed to look like. The pale bars are the forward question and they climb too, for a reason that has nothing to do with forecasting, since a month already inside a recession usually has more recession months behind it.
Read the first number in that table slowly, because it is the strongest claim on the page. Of the 613 months this model has read under 10%, NBER has dated 0 of them as part of a recession, which is not a small share but an empty one. At the other end, 62 of the 64 months reading 40% or more were recession months, and the 2 that were not all fall within 3 months of one starting, which makes them early rather than wrong.
The episode column is what stops that becoming a bigger claim than it is. Those 64 elevated months came from 10 separate runs, so anyone reading the top row is reading 10 events rather than 64 independent observations. The 59 years of monthly data behind this page still contain only 8 recessions, and no amount of monthly resolution creates more of them.
The fifth column looks like a forecast and is not one. A reading under 10% was followed by a recession inside a year in 15% of cases, and the median gap was 7 months, which is roughly what you would get by knowing nothing and noticing that recessions happen. The high bands score well on it because a month deep inside a recession is usually followed by more of the same recession, so the forward column is mostly measuring the length of the episode the reading already sits in.
The last 11 months of the record carry no figure in that column at all. Their following year has not finished, so NBER cannot have dated it, and scoring them would mean counting an unobserved future as a clean miss. They are excluded from the fifth column and counted in the fourth, which is why the two month counts differ.
The advance-warning test
Can this model see a recession coming?
No, and one test settles it. Take the model's reading exactly twelve months before each of the 8 recessions in the record, then put all 8 of them on the page, where there is nowhere for an average to hide a bad case and nowhere for a good case to carry the rest.
12 months before
Jan 1970
0.16%Reached 79.6% once it had started
12 months before
Dec 1973
0.24%Reached 99.1% once it had started
12 months before
Feb 1980
1.62%Reached 98.3% once it had started
12 months before
Aug 1981
0.86%Reached 93.8% once it had started
12 months before
Aug 1990
0.72%Reached 61.5% once it had started
12 months before
Apr 2001
0.40%Reached 30.0% once it had started
12 months before
Jan 2008
0.76%Reached 99.3% once it had started
12 months before
Mar 2020
1.60%Reached 100.0% once it had started
Not one of those 8 readings reached 10%. The loudest was 1.62%, a year out from February 1980, and the quietest was 0.16%, a year out from January 1970. A reading at or below the loudest of them describes 79% of every month in the record, so none of the 8 would have looked unusual at the time.
Would you rather see the fees before the sales call? The match tool asks six questions about your account and shows custodians that take your plan type, with their fee schedules, before anyone phones you. Start the match tool.
The reason is structural rather than a flaw anyone should fix. Payroll employment, industrial production, real income and real sales are four measures of activity that has already been counted, so a model reading them can only turn once the damage is in the data. That is the same reason it is so accurate about the present, and the two properties cannot be separated. An instrument that waits for confirmation is right when it speaks and silent when you would most want to hear from it.
The forward-looking version of this question is the Treasury yield curve, which inverts well before a downturn rather than during one, and the New York Fed turns that spread into a probability twelve months out. It shares no input with the model on this page. Our yield curve tracker carries the leading indicator and this page carries the confirmation, and reading the pair in that order is the only sequence that makes either one useful.
The record
What the model said before all 8 recessions
One row per recession, with the model's reading a year out, six months out, in the last month of the expansion and in the first month of the downturn. Nothing in this table is typed. It is read out of the two series on every build, which is why the numbers in the third and fourth columns are worth more than any sentence claiming the same thing.
| Recession | Months | 12 months out | 6 months out | At the peak | First month | Peak reading | Crossed 50% |
|---|---|---|---|---|---|---|---|
| January 1970 to November 1970 | 11 | 0.16% | 0.58% | 44.86% | 54.22% | 79.6% in Oct 1970 | month 1 |
| December 1973 to March 1975 | 16 | 0.24% | 0.34% | 2.46% | 37.64% | 99.1% in Nov 1974 | month 8 |
| February 1980 to July 1980 | 6 | 1.62% | 5.00% | 10.84% | 43.02% | 98.3% in May 1980 | month 2 |
| August 1981 to November 1982 | 16 | 0.86% | 4.16% | 5.40% | 46.46% | 93.8% in Nov 1981 | month 2 |
| August 1990 to March 1991 | 8 | 0.72% | 0.60% | 9.04% | 16.28% | 61.5% in Nov 1990 | month 3 |
| April 2001 to November 2001 | 8 | 0.40% | 4.68% | 25.96% | 30.00% | 30.0% in Apr 2001 | never |
| January 2008 to June 2009 | 18 | 0.76% | 1.38% | 9.30% | 22.38% | 99.3% in Sep 2008 | month 3 |
| March 2020 to April 2020 | 2 | 1.60% | 1.68% | 0.40% | 100.00% | 100.0% in Mar 2020 | month 1 |
The peak column is the business cycle peak, meaning the last month of the expansion. The NBER indicator uses the trough method, so the first recession month is the month after the peak and the December 2007 peak appears here as a January 2008 start
The third column tells the same story 8 times over. Twelve months before every recession in this record the model read under 10%, and the highest of those 8 readings was 1.62%. Six months out the highest reading across all 8 was 5.00%. The model is not failing at its job when it does that, because its job is to read four series measuring activity that has already been counted rather than to anticipate anything.
The March 2020 row shows the mechanism at its most extreme. In February 2020, the last month of the expansion, the model read 0.40%. One month later it read 100.00%, a jump of 99.6 points, because payrolls and industrial production had collapsed inside a single month and the model reads exactly those. Anyone waiting for this number to move had no time at all to act on it.
How long after each recession began the model crossed 50%
Lag from the first recession month to the alarm
Blue bars are the recessions the model called in their first month. Red bars are the ones it reached 50% on only after they were underway, and 1 of the 8 never crossed at all and cannot be drawn here. The slowest was December 1973, which had been running 7 months before the model agreed.
Where it failed
When has the recession probability model been wrong?
A calibration table reports averages, and averages hide the individual failures worth knowing about. These are the two lists that matter: the model's highest reading inside each stretch NBER dated as an expansion, and the recessions it read through without ever raising its voice.
Highest reading in each expansion
The false alarms
Each row is one continuous stretch of months NBER dated as expansion, with the model's highest reading inside it. Grouping them this way matters, because scored month by month a single elevated quarter reads as three separate failures.
- December 1969 Inside the 31 months of expansion from June 1967 44.86%
- April 1979 Inside the 58 months of expansion from April 1975 32.48%
- March 2001 Inside the 120 months of expansion from April 1991 25.96%
- December 1982 Inside the 92 months of expansion from December 1982 13.98%
- September 2005 Inside the 73 months of expansion from December 2001 12.16%
- July 1981 Inside the 12 months of expansion from August 1980 5.40%
The highest of them is 44.86% in December 1969 , and a recession NBER did date began 1 month later, so the model was reading a downturn ahead of the committee rather than one that never came.
Recessions it read through
The misses
NBER has dated 85 months as recession months inside this record. The model read 50% or more in 52 of them, so it stayed quiet through 33 months of downturn that it was supposed to be detecting.
Weakest call
30.0%Its highest reading across the April 2001 recession, which ran 8 months
Lowest reading in a recession
11.54%February 1982, which NBER dates as a recession month
The April 2001 recession is the clean failure. It ran 8 months, the model never once reached 50%, and it topped out at 30.0% in April 2001. A reader watching this number through 2001 would have seen nothing that looked like a recession, during a recession.
Both failure modes point the same way. The model reads activity rather than expectations, so it goes quiet through a shallow downturn that never dents the four series it watches, and it gets loud early only when the damage is already showing up in payrolls. Neither behaviour is a defect in the model, which was built to date business cycles rather than to anticipate them. Both are defects in using it as a warning.
Method
How is the recession probability calculated?
This page joins two series under one rule. The probability itself comes from a published model and is not recomputed here, while the scoring against NBER is ours, and the scoring is the part worth checking, because the same data supports two very different-looking numbers depending on which question is asked of it.
Marcelle Chauvet and Jeremy Piger estimate the probability with a dynamic-factor Markov-switching model reading four monthly series. Those are nonfarm payroll employment, the index of industrial production, real personal income excluding transfer payments, and real manufacturing and trade sales. All four measure activity that has already happened and all four publish on a lag, which is why the newest reading available is July 2026 while the calendar is further on.
The scoring runs against the NBER indicator, which carries a 1 for every month the committee dates as part of a recession. It uses NBER's trough method, meaning the flag turns on the month after the business cycle peak and stays on through the trough month. That is why the table above shows the Great Recession starting in January 2008 rather than at the December 2007 peak, and both months are printed so neither reader is misled.
The fourth column of the calibration table asks whether NBER dated the same month the reading refers to. The fifth asks whether it dated any of the following twelve months, and it is computed only for months whose whole following year the committee has finished dating. 11 months at the end of the record fall outside that window and are scored on the first question alone, because counting an unobserved future as a clean miss would flatter every row in the table.
Every series is joined on a year-and-month key rather than on a position in an array. A monthly file can be missing a month, and an index-based twelve-month lookback then measures thirteen months for everything after the gap and prints it as a year. The test suite rebuilds the calibration a second way from hand-built fixtures and compares the two, so a regression of that shape fails rather than ships.
The forward-looking counterpart to this page is the Treasury yield curve, which the New York Fed turns into a probability twelve months ahead. It shares no input with this model, so the two are independent evidence rather than one number wearing two hats, and reading them together is more useful than reading either alone. Our yield curve tracker carries that record and our Sahm rule tracker carries the unemployment-based version, which turns faster than this one and later than the curve.
Limits
What this recession probability does not show
Five things, and the first two are large enough that nothing on this page should be used as a forecast of anything.
It is not a forecast, and there is no horizon in it
The number is the estimated probability that a recession was underway in the month it is dated to. It carries no view about next month and no view about next year. The record column above shows it reading under 10% twelve months before all 8 recessions, which is what a coincident indicator does rather than a failure of one.
It reads one family of inputs
Four monthly activity series drive the whole model. Nothing in it sees credit spreads, the yield curve, house prices, equity markets, or anything about the rest of the world. A downturn that arrives through a channel those four do not measure is a downturn this number will be late to.
The sample is 8 recessions, not 710 months
Every table here prints an episode count beside its month count for that reason. 64 elevated months came from 10 runs, and a statistic computed on 8 events carries wide error bars however many monthly rows sit behind it. Two of those 8 were driven by an oil embargo and a pandemic.
A probability is not a date
Even when the model is loud it says nothing about how long the recession will last or how deep it goes. The 8 in this record with a declared trough ran from 2 months to 18 months, and the reading in the first month gives no help separating those two.
Both series get revised
The four inputs are revised, so a probability published for a recent month can move afterwards, and the model is smoothed which means the latest months move most. The NBER indicator is worse in one specific way: its recent months read zero until the committee announces a peak, which it does long after the fact, so an absence of a declaration is not evidence of no recession.
It says nothing about gold
The two series here are activity and recession dating. Neither measures demand for gold, and the gold section on this page finds no usable gap between what gold did after a high reading and what it did after any month at all. Anything about gold on this page is a measurement of the past rather than a reason to buy.
Gold
Does a high reading mean gold rises?
No, or at least not by enough to notice. Every month reading 40% or more that carries a gold close at both ends of a twelve-month window goes into one column, every other month goes into the other, and the two columns come out close enough that the difference is inside the noise.
After a reading of 40% or more
+5.3%Median next 12 months, 64 months from 10 episodes, positive in 66%
After any month in the record
+6.2%Median next 12 months, 689 months, positive in 62%
The difference
0.9 ptsOn a sample of 10 episodes, against a spread running −34.8% to +44.4%
Gold's median return over the twelve months after a reading in each band
The gap is 0.9 percentage points, which is smaller than it sounds because the spread inside those 10 episodes is far wider than the distance between them. Gold's twelve-month return after an elevated reading ran from −34.8% to +44.4% across the record, so the median hides two entirely different outcomes rather than describing a typical one.
The costs do not depend on the model being right about anything. A gold IRA charges a setup fee, an annual custodian fee and a depository storage fee that a brokerage account does not, and on top of those a dealer premium over spot running between 5% and 12% on IRA-eligible coins. On a $50,000 purchase the difference between those two premiums is $3,500 before the account has done anything, which is larger than the entire measured gap above. Our gold IRA fees page prices each of those separately.
Which form of gold you hold changes the cost more than any reading on this page does. A gold ETF and a mining stock both sit inside an ordinary brokerage IRA at brokerage cost, while bullion needs a self-directed IRA, a custodian and a depository, and the metal has to clear the IRS purity rule of 99.5% before a custodian will accept it into the account. Every figure here prices the metal at the London fix, so an ETF holder tracks that line minus an expense ratio and a mining stock does not track it at all. Our gold IRA rollover guide covers the account mechanics that sit underneath all of it.
How this page is built
Every figure is computed at build time from three committed data files, so the numbers are in the HTML before any script runs and every chart works with JavaScript turned off. The probability series carries 710 monthly readings from June 1967 to July 2026. The recession dating carries the whole NBER record. The gold series is the London fix at each month end from April 1968.
Nothing on this page is a live figure and nothing is dated to a month the underlying data does not reach. Corrections go to the editorial team and are logged publicly.
Sources
- FRED series RECPROUSM156N, Chauvet and Piger smoothed US recession probabilities
- FRED series USREC, the NBER based recession indicator
- NBER business cycle dating committee, US expansions and contractions
- LBMA precious metal prices, monthly gold fixes
- FRED series T10Y2Y, the 10-year minus 2-year Treasury spread behind the forward-looking model
We last checked these sources on 2026-09-09, and every figure on this page was rebuilt from them that day. Nothing here is typed by hand.
Full dataset available as CSV, with the probability, the NBER flag and the gold close in separate columns, so anyone who prefers a different band or a different horizon can rebuild the calibration themselves.
The Digest Perspective
Where we think this leaves you
· Written by the Gold IRA Digest Editorial Team · Reading the July 2026 print
The model read 0.76% for July 2026, its 75th consecutive month under 10%.
What the table above cannot settle is whether that silence carries information. A model with a perfect record of never speaking early is indistinguishable, from where we sit in July 2026, from one about to stay silent again, since both print 0.76%.
Chauvet and Piger
They built a business cycle chronology rather than a warning system and have never claimed otherwise, which is why the thing it does brilliantly is describe a month that has already been counted.
The NBER dating committee
They would rather be right than fast, so a peak is declared a year or more after it happened, which means the column this page scores against is itself unavailable in real time. Nobody in February 2020 knew they were living through the last month of an expansion.
The sales floor
A dealer needs this number to mean something in both directions, because low reads as complacency and high reads as urgency and either one closes a call, when in fact it describes a month that has already finished.
Where we are probably wrong
The strongest case against us runs the opposite way from the one we have made. An instrument right about 62 of 64 alarms, and never once wrong across 613 quiet months, has a cleaner record than any forecast we know of, and a reader could reasonably decide that knowing exactly where you are beats a guess about where you are going.
Our read: the best-calibrated instrument on this site is pointed at the wrong question for anyone making a decision. Use it for what has already happened, and use the curve and your own timeline for what might.
Which leaves the question that is actually yours. Not whether a recession is coming, since 8 readings above say nobody gets that a year ahead, but whether your account is positioned for one arriving at a time nobody can name, which depends on how close you are to withdrawals and on what the account already holds.
How much of your retirement is in one asset class?
The match tool asks six questions, then shows custodians that accept your plan type and your balance. You see the fee schedules before anyone calls you.
Six questions, about 2 minutes. We earn a fee when a custodian we match you with opens an account, and it does not change what you pay.
Common questions
Recession probability questions
- What is the current recession probability?
- 0.76% for July 2026, which is the most recent month Chauvet and Piger's model covers. That is the estimated probability the US economy was in a recession during that month, not the probability one arrives later. It is the 75 months in a row under 10%, and 86.3% of the 710 months since June 1967 sit in the same band, so a low reading here is the ordinary state of the series.
- How accurate is the recession probability model?
- Almost perfectly accurate about the present and useless about the future, and the calibration table above separates the two. NBER has dated 0 of the 613 months that read under 10% as recession months, and 62 of the 64 months that read 40% or more. The forward record is the opposite. Twelve months before each of the 8 recessions in the series the model read under 10%, and the highest of those readings was 1.62%.
- Does a high recession probability mean gold will rise?
- The record does not support that. After the 64 months reading 40% or more that carry a gold close at both ends, gold's median return over the following twelve months was +5.3% and it was positive in 66% of them. Across every month the two series share, the median is +6.2% and the positive share is 62%. Those 64 months also came from only 10 episodes, so the honest sample is 10 events.
- Is this the same as the yield curve recession model?
- No, and the difference is the reason to read both. This model reads four monthly measures of activity that has already happened, which are payroll employment, industrial production, real personal income excluding transfers, and real manufacturing and trade sales. The New York Fed's model reads the Treasury yield curve and estimates a probability twelve months out. They share no input, so they are independent evidence rather than two views of the same number. Our yield curve tracker covers the forward-looking one.
- Why does the model say zero when people are worried about a recession?
- Because it is not measuring worry, and it is not measuring the future either. It reads payrolls, industrial production, income and sales, so it turns only once those four have already broken. In February 2020 it read 0.40%, and one month later, in the first month NBER dates as recession, it read 100.00%. A reading of 0.76% says the four series were still intact in July 2026, and says nothing at all about January 2027.
- How often does this page update?
- The probability series, the NBER indicator and the gold series are rebuilt together on a schedule, and every figure here is computed at build time from those files rather than written by hand. The model runs about two months behind the calendar because the four activity series it reads publish on a lag, so July 2026 is the newest month available. Readings for recent months are revised as the underlying data is revised, and the section on limits above says what that means for anything quoted from here.
Gold IRA Digest is an independent publisher. This page is information, not investment advice, and nothing on it is a recommendation to buy or sell gold or any other asset. Past performance does not predict future results. Some links on this site earn us a commission, which never changes what we publish about a company.