Track Record — Cliff Radar Capital

§ 01

Fifty years, six recessions

Backtest · 1976–2026

Six recessions in fifty years. The machine was already defensive for five of them.

Our macro engine scores the whole economy into one number, and that number commands an 8-rung ladder — how much risk the models are allowed to carry. Here it is recomputed across 604 months of history, laid against the S&P 500. Shaded columns are the market's real peak-to-trough decline around each recession. One question: what was the machine saying while the market was topping — and what would listening have saved you?

−50% Losses AvoidedTHE 2008 CRASH: −57% BECAME −7%

In 2008 the S&P 500 lost 57% in one of the worst recessions on record. The EMP engine saw the stress building and stepped the models down — Take Profits at the October top, walked to Full Defensive before Lehman. A portfolio following that prescribed exposure would have fallen just 7%: a 50-percentage-point swing. And it isn't a one-off — across six recessions since 1980, an average 34% market decline came out as 11% at the machine's own exposure. Hypothetical model performance.

S&P 500 (log)EMP health scoreMarket peak → troughZones: green = risk-on rungs · amber = caution · red = defensiveNumbers mark each market peak ↓

1980 · market peak 1980-02-13 · panel coverage 58%

The 1980 Recession

At the top, the machine said
Full Defensive · 37.4%→ models 10% invested

Volcker's rate shock. Before the market turned, EMP was already at Full Defensive — a tenth invested. The decline happened almost entirely without you.

S&P 500 fell-17.1%

Following EMP-1.8%

Avoided15.3 pts

1981–82 · market peak 1980-11-28 · panel coverage 58%

The 1981–82 Recession

At the top, the machine said
Full Defensive · 42.2%→ models 10% invested

The double-dip. Full Defensive at the peak and defensive for twenty months. A 27% index decline came out as a 2% scratch.

S&P 500 fell-27.1%

Following EMP-2.2%

Avoided24.9 pts

1990–91 · market peak 1990-07-16 · panel coverage 64%

The 1990–91 Recession

At the top, the machine said
Sell Alert · 53.3%→ models 25% invested

Oil shock and credit stress. Sell Alert at the top — a quarter invested — so the drawdown landed on a book already mostly in cash.

S&P 500 fell-19.9%

Following EMP-4.0%

Avoided15.9 pts

2000–02 · market peak 2000-03-24 · panel coverage 72%

The Dot-Com Bust

At the top, the machine said
Sell Alert · 48.4%→ models 25% invested

Sell Alert in March 2000 while the crowd bought the dip — and nearly perfect through March 2002. Then the recovery gate began flipping monthly: a credit-data gap had disarmed the guard that should have stopped it.

S&P 500 fell-49.1%

Following EMP-22.2%

Avoided26.9 pts

2007–09 · market peak 2007-10-09 · panel coverage 79%

The Great Financial Crisis

At the top, the machine said
Take Profits · 63.8%→ models 70% invested

The best call in the record. Take Profits at the October 2007 top, walking down to Full Defensive before Lehman. The index lost 57%. The machine's path lost 7%.

S&P 500 fell-56.8%

Following EMP-7.3%

Avoided49.5 pts

2020 · market peak 2020-02-19 · panel coverage 86%

The COVID Crash

At the top, the machine said
Build Risk · 67.4%→ models 100% invested

The honest one. At the February 19 peak EMP said Build Risk — fully invested. A pandemic is not in macro data. It cut hard afterward, but it reacted; it did not warn. You would have felt nearly the whole crash.

S&P 500 fell-33.9%

Following EMP-27.6%

Avoided6.3 pts

Get today's regime read → See the live record since May 2026 →

§ 01·B

Fifty-five years, eight recessions

Backtest · 1970–2026

Eight recessions in fifty-five years. The model was above 50% before seven of them.

Alongside the EMP regime engine above, our recession engine collapses the credit cycle, the yield curve, the labor market and the housing pipeline into one number: the probability of a U.S. (NBER) recession beginning within twelve months. Here it is recomputed across every month since 1968, laid against the S&P 500. Shaded columns are the NBER recessions. One question: what was the model saying while the market was still climbing — and how much warning would listening have bought you?

7 of 8 Recessions PredictedRECESSIONS FLAGGED ABOVE 50% BEFORE ONSET · AVG PEAK 89%

Our recession-probability panel flagged 7 of the last 8 U.S. recessions before they began — its probability pushed above the 50% line a median of ten months ahead of onset, at an average peak of 89%. The lone miss was 2020: a pandemic is invisible to macro data, and the model topped out at 48% the month before lockdowns.

S&P 500 · log scale · monthly closes Recession probability · monthly 50% line NBER recession 50% warning fired (on the S&P) Numbered recession onsets

Probability is the blended 12-month model recomputed monthly with publication lags enforced — the number you could actually have seen in real time. The 2022–24 hump is a known false alarm from the sustained curve inversion (disclosed in full in the methodology at the bottom of this page).

1969–70 · Onset Jan 1970 · Crossed 50% Nov 1969

The 1969–70 Credit Crunch

In the year before, the model said
Peaked 59% · Crossed 50% Nov 1969

The model climbed out of the teens through late 1969 and breached 50% in November — two months before the downturn began — dipping briefly in December before surging as the recession took hold. A tight but correct call at the very start of the data.

Peak probability59%

Crossed 50%Nov 1969

Lead2 mo

1973–75 · Onset Nov 1973 · Crossed 50% Jul 1973

The Oil-Shock Recession

In the year before, the model said
Peaked 94% · Crossed 50% Jul 1973

Flat near 2% through the spring, then a vertical ramp as the curve inverted and permits rolled over: 65% by July, 85% by August, 94% by October — four months of warning before the November 1973 onset.

Peak probability94%

Crossed 50%Jul 1973

Lead4 mo

1980 · Onset Jan 1980 · Crossed 50% Mar 1979

The 1980 Credit Controls

In the year before, the model said
Peaked 99% · Crossed 50% Mar 1979

Volcker's first squeeze. The probability was above 50% from March 1979 and pinned near 99% for months before the January 1980 onset — ten months of warning.

Peak probability99%

Crossed 50%Mar 1979

Lead10 mo

1981–82 · Onset Jul 1981 · Crossed 50% Sep 1980

The Volcker Double-Dip

In the year before, the model said
Peaked 98% · Crossed 50% Sep 1980

The messy one. The model stayed 90–100% through the 1980 recession, cooled in the gap between the dips, re-fired to 66% in September 1980 and 98% by November, eased into the 20s–30s that spring, then jumped back to 80% in July 1981 — the month the second dip began.

Peak probability98%

Crossed 50%Sep 1980

Lead10 mo

1990–91 · Onset Jul 1990 · Crossed 50% Apr 1989

The Gulf-War Recession

In the year before, the model said
Peaked 78% · Crossed 50% Apr 1989

Flagged early — above 50% from April 1989 into early 1990 — then dipped in the spring before re-surging to 78% in June 1990, the month before the oil-price spike tipped the economy over.

Peak probability78%

Crossed 50%Apr 1989

Lead15 mo

2001 · Onset Mar 2001 · Crossed 50% Apr 2000

The Dot-Com Recession

In the year before, the model said
Peaked 96% · Crossed 50% Apr 2000

Crossed 50% in April 2000 as the curve inverted and hit 96% by that November — eleven months before the recession began in March 2001.

Peak probability96%

Crossed 50%Apr 2000

Lead11 mo

2008–09 · Onset Dec 2007 · Crossed 50% Nov 2006

The Global Financial Crisis

In the year before, the model said
Peaked 97% · Crossed 50% Nov 2006

First crossed 50% in November 2006, then from January 2007 held 85–97% through the entire year — thirteen months of elevated warning before the December 2007 onset of the deepest recession in the sample.

Peak probability97%

Crossed 50%Nov 2006

Lead13 mo

2020 · Onset Feb 2020 · Never Crossed 50%

The COVID Shock — the miss

In the year before, the model said
Peaked 48% · Never Crossed 50%

The one it could not see. Macro inputs were benign into 2020; the probability peaked at 48% in February, the month before lockdowns. No model built on slow-moving economic data can forecast a pandemic — shown here in full candor.

Peak probability48%

Crossed 50%Never

LeadNone

Get today's reading → See it beside the live record →

Ledger 1 · The Regime Ledger Free — always

Every macro call, dated. The weather report we actually printed.

The EMP health score, the band it commands on the public ladder, the tripwires lit, and the risk budget the models were ordered to run — beside the S&P's closing price, so you can scroll the tape against the calls. The receipts are free. Today's read lands with members before the open.

47 runs printed10 band changescurrent band Build Risk · held 1 session

Run date	EMP %	Regime band	Tripwires	Risk / Cash	S&P 500 close
August 2026
Aug 21, 2026	70.0%	Build Risk	1/11	90% / 10%	7,674
Aug 20, 2026	65.8%	Stay Invested	1/11	85% / 15%	7,641
Aug 19, 2026	65.9%	Stay Invested	1/11	85% / 15%	7,708
Aug 18, 2026	65.9%	Stay Invested	1/11	85% / 15%	7,692
Aug 17, 2026	65.9%	Stay Invested	1/11	85% / 15%	7,745
Aug 14, 2026	64.7%	Take Profits	1/11	85% / 15%	7,786
Aug 13, 2026	64.3%	Take Profits	1/11	85% / 15%	7,799
Aug 12, 2026	64.1%	Take Profits	1/11	85% / 15%	7,748
Aug 11, 2026	64.2%	Take Profits	1/11	85% / 15%	7,728
Aug 10, 2026	63.7%	Take Profits	1/11	85% / 15%	7,753
Aug 07, 2026	63.5%	Take Profits	1/11	85% / 15%	7,758
Aug 06, 2026	67.3%	Build Risk	1/11	90% / 10%	7,710
Aug 05, 2026	67.2%	Build Risk	1/11	90% / 10%	7,724
Aug 04, 2026	63.5%	Take Profits	1/11	85% / 15%	7,737
Aug 03, 2026	62.9%	Take Profits	1/11	85% / 15%	7,600
July 2026
Jul 31, 2026	66.9%	Stay Invested	1/11	91% / 9%	7,490
Jul 30, 2026	65.3%	Take Profits	1/11	85% / 15%	7,438
Jul 29, 2026	65.6%	Stay Invested	1/11	91% / 9%	7,316
Jul 28, 2026	65.5%	Stay Invested	1/11	91% / 9%	7,429
Jul 27, 2026	65.6%	Stay Invested	1/11	91% / 9%	7,413
Jul 24, 2026	65.6%	Stay Invested	1/11	91% / 9%	7,412
Jul 23, 2026	65.6%	Stay Invested	1/11	91% / 9%	7,408
Jul 22, 2026	65.6%	Stay Invested	1/11	91% / 9%	7,499
Jul 21, 2026	65.3%	Take Profits	1/11	85% / 15%	7,509
Jul 20, 2026	65.2%	Take Profits	1/11	85% / 15%	7,443
Jul 17, 2026	66.1%	Stay Invested	1/11	91% / 9%	7,458
Jul 16, 2026	65.9%	Stay Invested	1/11	91% / 9%	7,534
Jul 15, 2026	64.6%	Take Profits	1/11	85% / 15%	7,572
Jul 14, 2026	62.6%	Take Profits	2/11	85% / 15%	7,544
Jul 13, 2026	62.4%	Take Profits	2/11	85% / 15%	7,515
Jul 10, 2026	62.4%	Take Profits	2/11	85% / 15%	7,575
Jul 09, 2026	62.4%	Take Profits	2/11	85% / 15%	7,544
Jul 08, 2026	62.4%	Take Profits	2/11	85% / 15%	7,483
Jul 07, 2026	62.4%	Take Profits	2/11	85% / 15%	7,504
June 2026
Jun 30, 2026	63.1%	Take Profits	2/11	88% / 12%	7,499
Jun 29, 2026	62.9%	Take Profits	2/11	89% / 11%	7,440
Jun 24, 2026	63.5%	Take Profits	2/11	90% / 10%	7,358
Jun 16, 2026	63.4%	Take Profits	2/11	90% / 10%	7,511
Jun 11, 2026	61.5%	Take Profits	3/11	90% / 10%	7,394
Jun 09, 2026	64.3%	Take Profits	2/11	90% / 10%	7,387
Jun 05, 2026	64.8%	Take Profits	2/11	77% / 23%	7,384
Jun 04, 2026	64.8%	Take Profits	2/11	77% / 23%	7,584
Jun 03, 2026	64.6%	Take Profits	2/11	77% / 23%	7,554
Jun 01, 2026	64.3%	Take Profits	2/11	77% / 23%	7,600
May 2026
May 28, 2026	64.0%	Take Profits	2/11	77% / 23%	7,564
May 26, 2026	64.1%	Take Profits	2/11	77% / 23%	7,519
May 22, 2026	63.1%	Take Profits	2/11	77% / 23%	7,473

Aug 24, 2026’s read is members-first. Members receive it before the open; it posts to this public ledger the next trading day.

On a phone. Tripwires, Risk / Cash and the S&P close are hidden in portrait so the dates and bands fit. Turn the phone sideways for the full ledger — nothing has been removed from the record.

Reading the tape. Rows are grouped by month, and the tinted rows with the gold edge mark the sessions the regime band actually changed — the days the call moved are the story; everything between them is the machine holding its ground.

Coverage. This ledger prints every trading day since the desk went live on July 27, 2026. Rows before that date come from our build-and-test period, when runs were made as the system was being developed — they are shown exactly as they were printed, which is why early coverage is intermittent, and they are never backfilled. Going forward, publication may occasionally be delayed or missed for reasons outside our control (data-vendor, market-data or infrastructure outages among them); any resulting gap simply remains a gap, because this record is never edited or reconstructed after the fact.

Band shown is derived from the printed EMP score via the public 8-rung ladder on The System — deterministic and checkable from this table alone.

Focus Stock Setups

Every stock we named, and what it did next.

One row per dated Focus call — the price it was struck at, where it trades now, and the return at each milestone as it matures. Milestones are written once, when the call comes of age, and never rewritten. Days the machine named nothing are counted in the summary below, never dropped from the denominator.

Both horizons, and the days we said nothing.

7-day forward

+0.58%

n = 32 matured · 53% positive

every Focus call, 7 trading-day mark

30-day forward

+7.68%

n = 23 matured · 70% positive

every Focus call, 30-day mark

NO CALL days

31%

17 of 54 runs

no name cleared the bar — by design

Both horizons are published because both are the record. The 7-day series is the harder read and it stays on the page; the 30-day series is the one the model is built around. NO CALL days are counted as what they are — days the machine found nothing worth your money — never quietly dropped from the denominator.

Call date	Name	Struck at	Now	Live	7D	30D	60D	90D
Aug 24, 2026	QLYS Qualys Inc.	$181.98	$181.98	+0.0%	in 7d	in 30d	in 60d	in 90d
Aug 23, 2026	PBF PBF Energy	$73.53	$73.53	+0.0%	in 6d	in 29d	in 59d	in 89d
Aug 21, 2026	CHYM Chime Financial, Inc. Clas	$31.95	$32.78	+2.6%	in 4d	in 27d	in 57d	in 87d
Aug 20, 2026	NTAP NetApp Inc.	$194.48	$192.27	-1.1%	in 3d	in 26d	in 56d	in 86d
Aug 19, 2026	CRL Charles River Laboratories	$279.19	$295.19	+5.7%	in 2d	in 25d	in 55d	in 85d
Aug 17, 2026	GH Guardant Health	$157.21	$170.69	+8.6%	+5.6% ᶠ	in 23d	in 53d	in 83d
Aug 16, 2026	HNGE Hinge Health, Inc.	$88.49	$88.42	-0.1%	-3.2%	in 22d	in 52d	in 82d
Aug 14, 2026	GH Guardant Health	$159.54	$170.69	+7.0%	+2.4%	in 20d	in 50d	in 80d
Aug 13, 2026	AVTR Avantor	$13.93	$14.46	+3.8%	+0.2%	in 19d	in 49d	in 79d
Aug 12, 2026	FROG JFrog	$86.10	$92.13	+7.0%	+11.7%	in 18d	in 48d	in 78d
Aug 07, 2026	GH Guardant Health	$157.58	$170.69	+8.3%	+1.2%	in 13d	in 43d	in 73d
Jul 30, 2026	FTNT Fortinet Inc.	$160.08	$153.51	-4.1%	+9.1% ᶠ	in 5d	in 35d	in 65d
Jul 29, 2026	NTAP NetApp Inc.	$173.11	$192.27	+11.1%	+10.0%	in 4d	in 34d	in 64d
Jul 28, 2026	CVLT Commvault Systems	$149.46	$133.81	-10.5%	-16.4%	in 3d	in 33d	in 63d
Jul 24, 2026	GH Guardant Health	$150.20	$170.69	+13.6%	+11.5%	+8.7%	in 29d	in 59d
Jul 23, 2026	GH Guardant Health	$150.20	$170.69	+13.6%	+0.7%	+8.7%	in 28d	in 58d
Jul 22, 2026	OSCR Oscar Health	$30.77	$32.04	+4.1%	+1.1%	+2.6%	in 27d	in 57d
Jul 21, 2026	FROG JFrog	$90.43	$92.13	+1.9%	-10.2%	+6.3%	in 26d	in 56d
Jul 20, 2026	CRWD CrowdStrike Holdings	$203.96	$191.95	-5.9%	-10.1%	+4.4%	in 25d	in 55d
Jul 19, 2026	CNC Centene	$66.44	$65.02	-2.1%	-4.6%	+1.6%	in 24d	in 54d
Jul 17, 2026	OKTA Okta Inc.	$147.74	$135.14	-8.5%	-7.9%	-0.2%	in 22d	in 52d
Jul 16, 2026	OSCR Oscar Health	$30.61	$32.04	+4.7%	-3.6%	+0.6%	in 21d	in 51d
Jul 15, 2026	CVLT Commvault Systems	$145.16	$133.81	-7.8%	+0.9%	+4.1%	in 20d	in 50d
Jul 14, 2026	TGTX TG Therapeutics	$55.01	$54.27	-1.3%	-1.9%	-10.0%	in 19d	in 49d
Jul 13, 2026	TWST Twist Bioscience	$91.75	$145.59	+58.7%	-2.0%	+34.2%	in 18d	in 48d
Jul 10, 2026	TWST Twist Bioscience	$89.90	$145.59	+61.9%	+1.5%	+38.9%	in 15d	in 45d
Jul 09, 2026	CNC Centene	$67.50	$65.02	-3.7%	-1.3%	-4.9%	in 14d	in 44d
Jul 08, 2026	CNC Centene	$67.08	$65.02	-3.1%	+1.8%	-4.3%	in 13d	in 43d
Jul 07, 2026	BRKR Bruker	$59.03	$59.56	+0.9%	-0.1%	-11.1%	in 12d	in 42d
Jun 24, 2026	STX Seagate Technology Holding	$1,020.99	$850.00	-16.7%	-3.2%	-10.5%	-16.7%	in 29d
Jun 09, 2026	UMC United Microelectronics	$20.00	$18.34	-8.3%	+10.3%	+24.8%	-4.0%	in 14d
Jun 05, 2026	WDC Western Digital	$533.60	$459.44	-13.9%	-1.7%	+23.3%	-1.2%	in 10d
Jun 04, 2026	ON ON Semiconductor	$129.13	$74.21	-42.5%	-10.2%	-27.8%	-36.8%	in 9d
Jun 03, 2026	PANW Palo Alto Networks	$279.61	$357.87	+28.0%	-6.9%	+20.9%	+18.7%	in 8d
May 24, 2026	WDC Western Digital	$484.28	$459.44	-5.1%	+14.6%	+32.9% ᶠ	+15.3%	-3.1%
May 23, 2026	STX Seagate Technology Holding	$812.73	$850.00	+4.6%	+9.5%	+28.4%	+9.7%	+4.6%
May 22, 2026	ON ON Semiconductor	$116.20	$74.21	-36.1%	+9.8%	+5.1%	-25.4%	-34.1%

How to read this. Struck at is the close the call was published against. Now is the latest close (Aug 21, 2026), and Live is the move since the call — it keeps changing. The milestone columns do not: each is fixed the day it matures, which is why an old call and a new one can be compared at all. A milestone still running shows the days left. Hypothetical; no commissions, taxes, spreads or slippage. Past performance does not guarantee future results.

Ledger 2 · Model Performance

Foundation, Pulse and Strike — marked to market against the S&P 500.

Everything below is the live performance of our Tactical Radar System portfolios — the same three books anyone can follow on Autopilot and mirror in their own brokerage. Each is indexed to 100 where NAV marking began, shown beside the S&P 500 they answer to. They sit on a risk ladder: Foundation (lower risk), Pulse (medium), and Strike (higher).

MEASURED SINCE JUL 20, 2026 · 32 DAYS · 25 MARKED RUNS · THROUGH AUG 21, 2026 · INDEX = 100 AT START

How each line is built. Pulse and Strike are marked to market from their permanent position logs — real closing prices, stops honored, every fill on the ledger. Foundation is marked to market from its own allocation log — the holdings and weights it carried each day, priced at the close, the same way as Pulse and Strike. All three update automatically with every pipeline run. Live NAV marking began July 7, 2026 (Model Improvement 1, when the books moved to once-a-month entries and a disciplined mid-month exit review), which is why the curves start there and grow from here.

Hypothetical model-portfolio performance, not client accounts. No commissions, taxes, spreads or slippage are modeled; results are indexed, not dollar returns you could have earned. We publish them anyway — drawdowns and all — because a risk desk that hides its own drawdown isn't one. Past performance does not guarantee future results.

FoundationLower risk

The long-term core — think of it the way you'd think about a retirement account. Foundation reads the EMP macro regime to decide how much of the book belongs in broad equities versus a cash sleeve, leaning in when the backdrop is healthy and stepping back when the score flags stress. Built for steady compounding and shallower drawdowns rather than the biggest number in a good year.

Follow Foundation on Autopilot →

PulseMedium risk

The balanced growth book — Tactical Radar Pulse. Pulse spreads across the leading themes the system favors, holding a diversified set of names inside the equity budget the regime allows. More growth-oriented than Foundation and more diversified than Strike: conviction with guardrails, for an investor who wants to push returns without concentrating into a handful of bets.

Follow Pulse on Autopilot →

StrikeHigher risk

The aggressive book — Tactical Radar Strike. Strike runs the system's highest-conviction names without a theme budget holding it back, so it concentrates where the signal is strongest. The widest range of outcomes of the three — the largest upside in strong tapes, the deepest drawdowns when leadership breaks — for an investor who can stomach the swings.

Follow Strike on Autopilot →

The Proving Ground

Champion vs. Challengers — the models above are defending their seats.

Behind the live books — and the call desk, the put lab, the swing desk and the scoring core itself — registered challengers run every pipeline day as shadow harnesses — same dated marks, frozen rules, pre-registered promotion gates, and control lanes that are built to lose. A challenger that beats the champion out of sample, past its gate, becomes the live model. We are never satisfied with the champion — that is the point.

Gate progress — resize lanes: 0/24 out-of-sample resize Mondays · selection lanes: 0/12 out-of-sample months

Watch the challengers →

Page last rebuilt August 24, 2026. Full methodology on The System.

§ A

Backtest methodology & disclosures

Applies to § 01 and § 01·B

§ 01 · The EMP drawdown backtest — how those numbers were produced. This is a backtest: the EMP engine recomputed monthly over 1976–2026 macro history. Every index figure is a real peak-to-trough decline measured on daily closing prices of the S&P 500 (the 2007–09 line, for example, runs from 1,565.15 on 2007-10-09 to 676.53 on 2009-03-09). The "following EMP" figure holds the index at exactly the equity exposure EMP's band prescribed, rebalanced monthly, with each month's return governed by the prior month's reading — so the simulation never acts on information it could not have had. It pays no commissions, taxes, spreads or slippage, and earns nothing on cash; real cash yielded double digits in 1980, so the cash assumption is conservative while the cost assumption is optimistic. Band thresholds were fit on part of this history and validated on a hold-out set, which means this chart shows the rules working on data that helped shape them — the single most important caveat here. This is engineering validation, hypothetical by nature, and categorically separate from our live Track Record, which began May 22, 2026, publishes every trading day, and can never be backfilled.

§ 01 · What the history could not tell us. A backtest is only ever as good as the data put into it, and ours is incomplete the further back it reaches. Panel coverage runs from about 50% of indicators in the 1970s to 90% today — each card above prints the coverage behind that specific call, and the 1980–82 readings rest on barely half the panel. Scores are normalized across whatever indicators actually exist in a given month; an indicator with no data is excluded rather than guessed, and never counted as a fired crisis tripwire. High-yield credit spreads are the starkest case: in April 2026 our data provider truncated that series to a rolling three-year window under its licensing terms, deleting nearly three decades of history. Every call above except the most recent was therefore scored with no credit-spread input at all — including 2007–09, a crisis that was fundamentally about credit. We disclose that rather than quietly substitute a proxy. Macro data is also revised and restated after the fact, and several early series are proxies — a figure available to us today was not always available in real time. The engine is also evolving: this run reflects the version scored across history, while the live model adds refinements no backtest can retroactively earn credit for. Treat everything above as an estimate of what the rules would have done, not a record of what anyone did.

§ 01·B · The recession-probability backtest — how it is built and how well it scores. The engine is a ridge-logistic blend of two lenses — a Long-History lens fit from 1968 (eight recessions, including the 1970s inflation type) and a Credit-Cycle lens fit from 1990 that adds bank lending standards (SLOOS). Inputs are the 10-year/3-month term spread, the Baa credit spread and its 3-month change, building-permit growth, a real-time Sahm-rule labor signal, and heavy-truck sales — all lagged to their true publication dates, so the backtest reflects information available at the time, not hindsight. Validation is expanding-window walk-forward, refit annually and scored only on months the model had not yet seen: AUROC 0.896 out-of-sample (1997–2025), Brier score 0.126, and a base-leg AUROC of 0.863 since 1980 against 0.685 for the yield curve alone. Seven of eight recessions crossed the 50% line before onset; the S&P series shown is monthly closes (via FMP), recession dating is NBER's.

§ 01·B · The false alarm, disclosed. Every curve-based recession model — including this one — produced a sustained false signal during the 2022–24 yield-curve inversion, the visible hump on the chart with no shaded recession behind it. That is precisely why the output is published as a probability rather than a binary siren, and why the 50% line is a reference, not a trading rule. The lenses were fit on parts of this same history, so the in-sample portions of the chart show the rules working on data that helped shape them. The current live reading — 17% as of the June 2026 data month — is produced by the same engine and updates as new data publishes.

The limits of both backtests, stated plainly. Results computed on history are estimates, not experience: no one earned these returns, no account traded this way, and nothing here is a promise, projection, or guarantee of any future result. A model that warned before past recessions can stay quiet before the next one — the 2020 cards in both sections show exactly that happening. Cliff Radar Capital, LLC is a publisher of general market research, not an investment adviser, broker-dealer, or fiduciary to anyone. Nothing on this page is investment advice, a recommendation, an offer or solicitation, or personalized to your circumstances, and no adviser–client relationship is created by reading it. Do not rely on it as the basis for any investment decision; consult a licensed professional who knows your situation. All investing involves risk, including the total loss of principal. To the fullest extent permitted by law, Cliff Radar Capital, LLC, its members and contributors disclaim all warranties, express or implied, as to the accuracy or completeness of this analysis and accept no liability for any loss or damage of any kind arising from its use. You use it at your own risk.

The macro call, dated. The models, marked to market.