Macro Regime And Recession Indicators

Macro regime and recession indicators

The NBER does not tell you that you are in a recession. It tells you, on average seven months later, that you were. Every recession indicator on FRED is scored against a label with that lag baked in, and the label itself is served as a clean monthly series with a value for every month of its long history, including the months nobody could have labelled.

../../../fin-core/skills/regime-detection/SKILL.md owns the estimators — HMMs, the Hamilton filter, smoothed-versus-filtered-versus-predicted probabilities. This skill owns the label and the three published indicators, and it never fits a regime model.

Everything marked ✅ Measured comes from scripts/recession_indicators.py (numpy + pandas + scipy, seed 20260909, 12 s, no network). §1 is arithmetic on the NBER's own published announcement dates; §§2–4 run on a synthetic economy whose recession lengths and announcement lags are the real US ones.

1. ✅ The announcement lag, from the committee's own table

✅ source-verified at nber.org/research/business-cycle-dating/business-cycle-dating-procedure-frequently-asked-questions, read 2026-09-09. The committee publishes the lag itself: "There is no fixed timing rule because the committee waits long enough to avoid any doubt about the existence of a peak or trough", and "The committee also allows sufficient time for standard data revisions in order to assign an accurate peak or trough date."

✅ Measured — the script recomputes every lag from the two dates and all 12 match the column NBER prints. Peaks: mean 7.3 months, range 4–12. Troughs: mean 15.2 months, range 8–21.

🚨 The December 2007 peak was announced on 2008-12-01. Lehman had already failed. And the April 2020 trough was announced on 2021-07-19 — fifteen months after the recovery began.

The number that follows

✅ Measured on those dates, using FRED's own USREC convention ("the recession begins the first day of the period following a peak and ends on the last day of the period of the trough"): a recession month is knowable only once the peak has been announced.

🚨 64 % of US recession months since 1980 carried no NBER recession label while they were happening. USREC has one for every single month. The 1990–91 recession was over before it was declared.

2. 🚨 The vintage label does not abstain — it disagrees

✅ source-verified at fred.stlouisfed.org/series/USREC: "This time series is an interpretation of US Business Cycle Expansions and Contractions data provided by The National Bureau of Economic Research (NBER)." An interpretation of a dated list — so the vintage USREC for an undeclared recession is not missing. It is a published, confident zero, and after the recession ends it stays at one until the trough is announced.

def usrec_known_at(state, peak_lag, trough_lag, d):
    out = np.zeros(state.size)                    # NOT NaN - that is the whole point
    for i, (s, e) in enumerate(episodes(state)):
        if s + peak_lag[i] > d:
            continue                              # peak not announced: this reads 0
        stop = e if e + trough_lag[i] <= d else min(d, state.size - 1)
        out[s:stop + 1] = 1.0                     # and it overstays at the other end
    out[d + 1:] = np.nan
    return out

The honest reconstruction is ALFRED's USREC vintages, or this lag applied to the dated list. The ../real-time-macro-backtesting/SKILL.md as_of() pattern is the same shape.

3. ✅ Measured — how much of "recession-aware" is the label

The synthetic economy uses the real US recession lengths since 1980 (6, 16, 8, 8, 18, 2 months) and the real NBER announcement lags, so its live-label statistics are comparable to §1's. 900 months, 9 recessions, 9.4 % of months in recession; everything is scored on the last 480.

✅ Measured — the same series, read two ways:

🚨 +0.389 of AUC between reading USREC retrospectively and reading it as of the month. Live, it catches 35 % of recession months at 23 % precision — which reproduces §1's real-data 36 %, because the synthetic cycle inherits the real lengths and lags. A perfect recession indicator is what USREC looks like from the future.

🔑 But a fitted classifier barely notices the label. ✅ Measured over 12 seeds:

The mislabelled rows are 0.17 % of an expanding training window (4.3 % of its last 24 months), so the fitted coefficients hardly move; the cost is two orders of magnitude below the 0.389 that reading USREC as the signal costs. At a 120-month window the comparison dissolves into noise and changes sign. 🔑 So the fix is not "retrain on vintage labels". It is: never use USREC as a live flag, and never evaluate against it as though it were one.

4. ✅ The Sahm rule, and what its own inputs do to it

✅ source-verified, fred.stlouisfed.org/series/SAHMREALTIME, verbatim: "Sahm Recession Indicator signals the start of a recession when the three-month moving average of the national unemployment rate (U3) rises by 0.50 percentage points or more relative to the minimum of the three-month averages from the previous 12 months."

🔑 FRED publishes the same rule twice. SAHMREALTIME adds: "This indicator is based on 'real-time' data, that is, the unemployment rate (and the recent history of unemployment rates) that were available in a given month." SAHMCURRENT's note is ✅ verified identical minus that paragraph — same rule, today's series. The two series exist because the answers differ, and only one of the two pages tells you which object you are holding.

✅ source-verified on the same page, the reason U3 was chosen: "The BLS revises the unemployment rate each year at the beginning of January… Revisions to the seasonal factors can affect estimates in recent years. Otherwise the unemployment rate does not revise." Compare payrolls, whose mean absolute first-to-third revision is 51k (../real-time-macro-backtesting/SKILL.md §1). The rule's robustness is a design choice about its input, not a property of thresholds.

✅ Measured, with annual seasonal-factor revisions of 0.075 pp reaching back five years:

Push the input revisions up and the hard threshold is what breaks — ✅ Measured:

🚨 The value moves by hundredths and the SIGNAL moves by months. A hard threshold converts a rounding-level input revision into a different trigger date and, past 0.15 pp, into false alarms out of nothing. Any indicator of the form x > c inherits this; report the distance to the threshold, not just the boolean.

5. Yield-curve inversion

✅ Measured on the synthetic economy, where the curve is built to invert 4–22 months ahead of each recession plus six inversions that lead to nothing — so the generative claim is explicit rather than discovered:

🔑 The lead is the problem, not the hit rate. A signal with an interquartile range of seven months is not a position size, and it is not a stop. And the only way to score whether a given inversion was a false alarm is the NBER label — which arrives 7.3 months after a peak and 15.2 after a trough. You cannot backtest the indicator faster than the referee reports.

⚠️ The lead distribution here is the script's own assumption, made visible. Do not read these as US historical lead times; read the mechanism.

6. Traps

• 🚨 USREC as a regime flag, a feature, or a target. §1 and §3. It is a retrospective interpretation of a dated list. If you need a live one, use ALFRED's USREC vintages, or apply the announcement lag to the dated list yourself.
• 🚨 RECPROUSM156N and the other "smoothed recession probability" series are model output, not observation — and they are estimated on the current vintage of their inputs. ⚠️ Whether each such series is available with ALFRED vintages is per-series; check before designing a study around it.
• 🚨 Reporting a boolean instead of a distance. §4: sahm >= 0.50 flips on a 0.02 pp wobble when the value sits at 0.49. Log the value, and log how far it is from the threshold.
• 🚨 SAHMCURRENT in a backtest. It is the rule computed on today's unemployment series. SAHMREALTIME is the one that existed. Nothing in the value tells you which you have.
• 🚨 Scoring a leading indicator against the label it leads. An inversion 14 months before a recession looks like a false alarm for 14 months, and then like genius — and both readings need a turning point that will not be dated for another year.
• 🚨 Fitting a regime model and calling the smoothed probability a signal. That is a different look-ahead and ../../../fin-core/skills/regime-detection/SKILL.md measures it: a full-sample fit anticipated 47 % of regime switches before they started. Both bugs can be in the same pipeline.
• ⚠️ The 2020 recession is two months long. Any indicator with a 3-month moving average cannot resolve it, and any train/test split that straddles it is dominated by it. State how you handled 2020 or the result is not comparable to anyone else's.
• 🚨 A recession indicator is a trial. Threshold, smoothing window, which spread, which unemployment measure. Log them in ../../../fin-core/skills/backtest-validation/scripts/trial_ledger.py.

❓ Not verified

Whether ALFRED carries per-observation vintages for USREC specifically (the announcement dates in §1 are the primary source either way) · the historical US inversion-to-recession lead distribution — §5's numbers are the script's own DGP, not history · RECPROUSM156N's estimation vintage.

7. Scripts and where this sits

scripts/recession_indicators.py — the NBER announcement table with its own self-check, the unlabelled-month computation, usrec_known_at() / live_usrec(), the classifier comparison and the 12-seed label-cost sweep, the Sahm rule with its vintage/revised comparison and threshold-tolerance sweep, and the inversion record. numpy + pandas + scipy, seed 20260909, 12 s, no network.

• HMMs, the Hamilton filter, and smoothed vs filtered vs predicted regime probabilities — ../../../fin-core/skills/regime-detection/SKILL.md. That skill owns the estimator; this one owns the label.
• The vintage A/B for a macro strategy, and as_of() — ../real-time-macro-backtesting/SKILL.md.
• Why unemployment's seasonal factors move at all — ../seasonal-adjustment-and-x13/SKILL.md.
• Nowcasting the quarter you are in — ../gdp-nowcasting-dynamic-factor/SKILL.md.
• When the underlying prints land — ../macro-release-calendar-and-embargo/SKILL.md.
• Where USREC, SAHMREALTIME and T10Y3M come from — ../../../fin-core/skills/fundamental-and-macro-data/SKILL.md.
• Whether a regime-conditional result survives its own sample — ../../../fin-core/skills/regime-detection/scripts/regime_coverage.py.