Backtesting evidence

Long-history: does the credit gate survive 55 years? (Jun 2026)

Versioned source evidence · docs/LONG_HISTORY_CREDIT_GATE.md

SHA-256: 84027485b02dc7c2b9a4616c08fa91284a72c13790ead6c9da8923e54126d509

Original source text. A superseded report retains its original assertions; it does not replace a corrected finding. Reproduction code is displayed without execution.

# Long-history: does the credit gate survive 55 years? (Jun 2026)

The in-app `stovall_credit_bonds` study runs ~20 years because the SPDR sector
ETFs only launched in Dec 1998 — **the limit is the equity data, not the
strategy.** The credit gate's own inputs go far deeper (`NFCICREDIT` to 1971,
`DGS10` to 1962), so we can keep the *exact* gate and feed it long-history
**index** data instead of ETFs, and watch it run through five decades of crises
the 20-year window never saw.

## TL;DR

> **Rewritten Sep 2026.** Every number above the fold used to be ~2pp of CAGR
> and ~0.25 of Sharpe too generous. The overlay carried three errors — exposure
> and the rate switch both stamped on the current week, and the cash leg booked
> at 0% — and this file's headline was the loudest surviving statement of the
> pre-correction story. It said the gate "roughly doubles the money while
> halving the drawdown, **with higher** total return". It does not. Corrected
> here and in `scripts/long_history_credit_gate.py` (PR #558 fixed the in-app
> module; the script itself was missed until now). The correction history is
> kept in full further down rather than quietly overwritten.

**The gate halves drawdown over 54.5 years (1971→2026) and pays for it in
return.** Applying the production overlay (exposure = `1 −
trailing-percentile(NFCICREDIT)`, de-risked fraction → intermediate Treasuries
when the 10y is falling, else T-bills), with both signals lagged to what was
actually knowable:

| leg (1971→2026, 54.5y) | CAGR | Sharpe | maxDD | Calmar | $1 → |
|---|---:|---:|---:|---:|---:|
| US market buy & hold | 11.2% | 0.72 | −54.0% | 0.21 | $336 |
| **Market + credit overlay** | **10.4%** | **1.17** | **−26.8%** | **0.39** | **$226** |
| 12-industry equal-weight basket | 11.5% | 0.76 | −53.5% | 0.21 | $377 |
| **12-ind EW + credit overlay** | **10.6%** | **1.21** | **−24.8%** | **0.43** | **$242** |

**It gives up ~0.9pp of CAGR to cut the drawdown roughly in half**, on *both* a
value-weighted-market base and an equal-weight-industry base, net of 10 bps/side.
Sharpe 0.72 → ~1.2 because volatility falls faster than return does. It ends
with **less money** than buy & hold ($226 vs $336) — the compounding cost of
being ~half invested is real and this is what it looks like over five decades.

Whether that trade is worth making is a risk-tolerance question, not a
back-test result. What the study establishes is that the drawdown reduction is
robust across five decades and two equity bases; it does not establish an edge.

This is the same lift seen across markets in
[`OUT_OF_SAMPLE_CREDIT_GATE.md`](OUT_OF_SAMPLE_CREDIT_GATE.md) — now confirmed
across **time** as well.

> **Corrected Sep 2026 (PR #561).** That companion study carried the same two
> errors corrected here (both signals unlagged, cash booked at 0%), and its
> Sharpe headline does not survive them: developed-market countries are
> **0.49 → 0.89** and European sectors **0.43 → 0.90**, not the 1.12/1.22 once
> published. The claim that the EU result "almost exactly reproduces" the US
> one is withdrawn — it is EU 0.90 against US 1.24. What survives, and what the
> cross-market case should rest on, is the **drawdown**: −58.7% → −20.7% on
> countries and −50.0% → −15.9% on EU sectors, in all four walk-forward halves
> and every crisis window.

## It improves risk-adjusted return in every decade

| decade | market Sharpe → +overlay | maxDD → +overlay |
|---|---|---|
| 1970s | 0.43 → **0.99** | −48.0% → **−17.2%** |
| 1980s | 1.07 → **1.43** | −32.9% → **−14.4%** |
| 1990s | 1.31 → **1.65** | −19.9% → **−7.4%** |
| 2000s | 0.09 → **0.75** | −54.0% → **−26.8%** |
| 2010s | 0.95 → **1.01** | −18.3% → **−13.8%** |
| 2020s | 0.82 → **1.22** | −32.2% → **−13.4%** |

Six decades, no exception — including the *lost decade* of the 2000s, where the
gate turned a 0.09-Sharpe market into 0.75. This is the claim that survived the
correction most intact; the 2010s is the thinnest (0.95 → 1.01), which is the
decade with no credit event in it.

## It de-risked every crisis in five decades (market base → +overlay)

| crisis | market: return / maxDD | + overlay: return / maxDD |
|---|---|---|
| 1973–74 stagflation | −42.9% / −48.0% | **−1.1% / −17.2%** |
| Volcker 1980–82 | +1.3% / −19.0% | **+20.9% / −12.6%** |
| Black Monday 1987 | −20.5% / −32.9% | **−9.8% / −14.4%** |
| 1990 recession | −15.4% / −19.4% | **−0.5% / −2.0%** |
| Dot-com 2000–02 | −34.4% / −47.5% | **−11.6% / −26.4%** |
| GFC 2008 | −43.3% / −54.0% | **+14.3% / −6.9%** |
| COVID 2020 | −11.9% / −32.2% | **+8.2% / −2.5%** |
| Rate-bear 2022 | −18.3% / −23.1% | **−5.8% / −5.4%** |

The gate was **net positive** through Volcker, 1990, dot-com, GFC and COVID, and
turned the rest into single-digit-ish drawdowns. The rate-regime switch is what
makes it work in the **inflation** eras (1973–74, 2022): rates *rising* → the
de-risked sleeve sits in **cash**, not bonds, so it doesn't compound the
duration loss. 1973–74 is the overlay's weakest crisis (−16.9%) — rising rates
forced cash, and `NFCICREDIT`'s trailing percentile was still warming up that
early — but it still beat the market's −48% by three-to-one.

## What the gate does and does not see

Average exposure through each window is not a footnote — it is the mechanism,
and it predicts where the protection shows up. Sorted by how invested the gate
stayed:

| window | mean equity held | benchmark drawdown | credit event? |
|---|---:|---:|---|
| Tariff shock 2025 | 77% | −17.7% | no |
| EU debt 2011 | 67% | −18.3% | no (sovereign, not US credit) |
| Q4 2018 | 61% | −18.3% | no (liquidity/hiking) |
| Black Monday 1987 | 39% | −32.9% | no |
| Dot-com 2000–02 | 31% | −47.5% | no (valuation de-rating) |
| 1973–74 stagflation | 31% | −48.0% | no (inflation shock) |
| COVID 2020 | 23% | −32.2% | partly |
| Rate-bear 2022 | 21% | −24.9% | partly |
| Volcker 1980–82 | 16% | −19.0% | partly |
| 1990 recession | 4% | −19.4% | **yes** (S&L bust) |
| GFC 2007–09 | 4% | −54.0% | **yes** |
| China / oil 2015–16 | 1% | −13.7% | **yes** (HY energy) |
| LTCM / Asia 1998 | 0% | −19.9% | **yes** |

**The gate protects in proportion to how much the crisis is a credit event.**
That is not a caveat, it is the design: exposure is `1 − percentile(NFCICREDIT)`,
so a drawdown that never reaches the credit market is close to invisible to it.
Dot-com is the clearest case — the deepest non-credit drawdown in the sample,
and the gate rode 31% of it.

The 20-year production study reaches the same number from a completely different
equity base: 32% average exposure through dot-com against 7% in the GFC (mine:
31% and 4%). Two bases, one mechanism.

**Practical reading:** do not expect this to help in a valuation de-rating, an
inflation shock, or a fast liquidity air-pocket. Expect it to help when spreads
widen. The 2010s being the thinnest decade in the table above is the same fact
seen from the other side — no credit event, no lift.

## Method

- **Equity (index, not ETF):** Fama-French daily series (1926+) — the broad
  market (`Mkt = Mkt-RF + RF`) and the 12 industry portfolios (a real,
  deep-history sector universe). Daily simple returns compounded to W-FRI weekly.
- **Gate:** imported `_credit_exposure_series` (NFCICREDIT exposure) and
  `CREDIT_RATE_MA_WEEKS` (DGS10 rate regime) from
  [`rotation_strategies.py`](../backend/app/scoring/rotation_strategies.py) —
  *verbatim*, same constants the live signal uses.
- **Defensive sleeve:** intermediate-Treasury total return = **actual IEF** from
  2002, spliced over a **DGS10-constructed** return before that (modified
  duration D≈7.5). The construction was validated against real IEF over 2002–26:
  **0.98 weekly correlation, matched annualised vol**, and it slightly
  *understates* bond return — so the deep-history leg is **conservative**.
- Effective span 1971-12 → 2026-05 (~54.5y) once the 156-week credit percentile
  warms up. Net of 10 bps/side.

## Caveats

- **The overlay, not the full Stovall strategy.** Stovall sector *selection* is
  ETF-era and US-cycle-specific; it can't be run pre-1998. This tests the
  *transferable* half — the credit gate — which the research identified as the
  actual edge. The equity base here is the broad market / EW industries, not
  Stovall's leader sectors, so *absolute* returns aren't the live strategy's;
  the **lift from the gate** is the result.
- **Constructed bonds pre-2002.** Validated and conservative (above), but it is a
  model, not a traded series, for the 1971–2002 portion of the defensive sleeve.
- **Fama-French industries ≠ SPDR sectors** (different classification, weighting)
  — a feature here (deep history, and shows the result isn't SPDR-specific).
- **One country (US).** Breadth across *markets* is the companion study
  ([`OUT_OF_SAMPLE_CREDIT_GATE.md`](OUT_OF_SAMPLE_CREDIT_GATE.md)); this is the
  *depth-in-time* axis.

## In the app: the crisis-by-crisis chart (Sep 2026)

The study above lived only as a printed table. It now ships as a chart —
`/backtest/stovall`, backed by `GET /backtest/credit-gate/crisis-periods`
([`credit_gate_crisis.py`](../backend/app/scoring/credit_gate_crisis.py)),
which re-runs this exact overlay in-app. Three deliberate differences from
the script that produced the table above:

**1. Signal timing — and a correction to this whole study.** The script stamps
*both* of the gate's signals on the current week: it multiplies week *t*'s
exposure by week *t*'s own return, and routes the de-risked sleeve using week
*t*'s own rate regime. Both are one-week look-aheads, and the second one is not
small. `falling` is a **backward**-looking condition — the 10y under its 26-week
mean means bonds have *already* been rallying — so stamping it at *t* books the
sleeve straight into the rally the switch is measuring:
`E[bond return | falling@t]` = **+0.37%/wk** against **+0.18%/wk** once the
switch is lagged (unconditional +0.12%/wk).

The module lags both signals. That is the difference between a strategy that
appears to beat the market and one that does not:

| 1971→2026, 2845 weeks, 10bps/side | CAGR | Sharpe | maxDD | $1 → |
|---|---:|---:|---:|---:|
| US market buy & hold | 11.2% | 0.72 | −54.0% | $336 |
| script convention (both signals same-week, 0% cash) | 12.7% | 1.41 | −19.3% | $709 |
| exposure lagged, rate switch not | 12.4% | 1.37 | −20.2% | $588 |
| both lagged, cash still booked at 0% | 9.3% | 1.06 | −26.6% | $131 |
| both lagged 1w, cash earns the T-bill | 10.9% | 1.22 | −25.9% | $280 |
| **shipped: NFCI lagged 2w (the measured release lag)** | **10.4%** | **1.17** | **−26.8%** | **$226** |
| *placebo: always hold bonds, no rate switch* | *10.3%* | *1.05* | *−22.9%* | *$211* |
| *placebo: always T-bills, no bond sleeve* | *9.5%* | *1.18* | *−24.4%* | *$140* |

The last two rows of the correction matter as much as the first. **Cash is not
0%.** The de-risked half sits in cash whenever rates are rising, and over a
window containing 15% T-bills, booking that at zero is a one-sided penalty worth
~1.5pp a year — the mirror image of the timing leak. The shipped module pays it
the one-month bill (Fama-French `RF`, already in the same file).

Net of all three corrections the honest read is: **a little behind buy & hold
on return (10.4% vs 11.2%), for half the drawdown (−26.8% vs −54.0%), Sharpe
0.72 → 1.17.** What is *not* supported is the original "roughly doubles the
money while halving the drawdown". The tables at the top of this file, the
decade table and the crisis table all used to carry the same-week stamping and
the 0% cash leg; they were regenerated from the corrected script in Sep 2026 and
now agree with the in-app module to the decimal. The figures preserved in this
section are the *superseded* ones, kept so the size and direction of the
correction stay auditable.

> **Do not read this as confirming the house "half the drawdown for about the
> same return" line** ([`NEGATIVE_FINDINGS`](NEGATIVE_FINDINGS.md) #7). An
> earlier draft of this section said it did. That framing was itself computed
> before the timing and cash-leg corrections, and PR #565 finds it does not
> survive them — the out-of-sample shortfall against SPY widens from −1.8pp to
> −4.1pp/yr, and #7's full-cycle rebuttal is now flagged UNVERIFIED pending a
> re-run. The −0.8pp shortfall measured here is this study's own number on a
> different base and window; it is not corroboration of #7, and #7 should not
> be cited as corroboration of it. Two studies computed with the same bug
> agreeing is not evidence — see the method notes in
> [`ROTATION_CREDIT_RESEARCH.md`](ROTATION_CREDIT_RESEARCH.md).

> **Drawdown seeding (Sep 2026).** This module computed its drawdowns from the
> NAV series *after* the first return, so a fall beginning at starting capital
> was invisible — the same seed bug fixed repo-wide in PR #564, reproduced here
> independently. It moves exactly one of the thirteen windows: **Rate-bear
> 2022**, which opens at the January peak, from −23.1% / −6.2% to
> **−24.9% / −7.6%** (benchmark / gate). Both legs understate equally, so no
> comparison flipped, and the full-history figures are unchanged because
> neither leg's worst drawdown starts on the first bar. Flagged by the session
> that shipped #564 — my sample has the shape the bug hides and theirs did not.

### Does it survive a publication lag? Only if the crisis is slow.

`NFCICREDIT` prints on Wednesday carrying data through the *prior* Friday, so a
one-week lag is optimistic about what you could actually have traded — which is
why the shipped convention is two. Sweeping it, on the fair (T-bill cash)
sleeve:

| NFCI lag | CAGR | Sharpe | maxDD |
|---|---:|---:|---:|
| 1 week (not actually knowable — see below) | 10.9% | 1.22 | −25.9% |
| **2 weeks (shipped: the measured release lag)** | **10.4%** | **1.17** | **−26.8%** |
| 3 weeks | 9.9% | 1.10 | −28.2% |
| 4 weeks | 9.5% | 1.04 | −28.7% |

**Two weeks is this module's offset, not a universal constant.** It is
wall-clock-correct for an engine whose weekly bar is the return over
`(w-1, w]`; the production engine stamps exposure at the *start* of the
interval it governs, so the same real-world timing is a shift of **1** there.
Copying the constant across would over-lag production by a week. Check the
indexing before reusing the number.

**The real lag is two weeks here, and it is measured, not assumed.** Two
independently-dated pulls of `NFCICREDIT` pin the release rule: fetched Tue
2026-06-23 its last observation was Fri 2026-06-12 (an 11-day gap — the day
before that week's release), and fetched Thu 2026-09-03 it was Fri 2026-08-28
(a 6-day gap — the day after). Both match **released Wednesday, covering the
week ending the prior Friday**. So on a W-FRI grid the print dated Friday *t−1*
does not exist until the Wednesday of week *t*, halfway through the week a
one-week shift applies it to. Two weeks is the first lag genuinely knowable
before the week opens, so that is what ships. `DGS10` needs only one week — it
is daily and published the same evening.

Roughly −0.45pp of CAGR per extra week, against a full-history drawdown that
barely moves. **That aggregate is misleading, and it took a second opinion to
catch it.** The full-history maximum drawdown is set by the slow grinds
(dot-com, 1973–74), so it is nearly lag-insensitive by construction — while the
protection against *fast* crashes falls apart. Per window, lag 1 → lag 4:

| crisis | wks to trough | bench dd/wk | lag 1 | lag 4 | degradation |
|---|---:|---:|---:|---:|---:|
| COVID 2020 | 7 | 4.6% | −2.5% | **−13.6%** | **+436%** |
| Black Monday 1987 | 18 | 1.8% | −12.9% | −17.2% | +34% |
| Tariff shock 2025 | 9 | 2.0% | −12.6% | −15.2% | +21% |
| Q4 2018 | 16 | 1.1% | −9.9% | −11.5% | +15% |
| … | | | | | |
| GFC 2007–09 | 75 | 0.7% | −6.6% | −7.2% | +10% |
| Dot-com 2000–02 | 136 | 0.4% | −25.5% | −28.6% | +12% |
| 1973–74 stagflation | 92 | 0.5% | −16.8% | −17.5% | +4% |
| Volcker 1980–82 | 93 | 0.2% | −12.4% | −12.7% | +2% |

**COVID surviving at the tradeable lag is luck, not robustness.** The real lag
lands at 2 weeks and the fast-crisis breakdown starts at 3 — two weeks of margin
we did nothing to earn. Had the Chicago Fed published a week later, the gate
would have failed the single fastest crisis in the sample. Read the survival as
"we got away with it", not as evidence the design is delay-tolerant.

**corr(crisis speed, lag degradation) = +0.90** across all 13 windows. The five
fast windows (≥1%/wk) degrade **+101%** on average; the eight slow ones **+14%**.
COVID is the extreme case and the instructive one: the gate entered that window
89% invested and de-risked *during* the crash, so a signal four weeks late
simply misses it.

So the honest statement is **not** "the drawdown protection is robust to signal
delay". It is: **robust against slow drawdowns, and it fails against fast ones —
and the fast ones are what the last two decades look like.** A reader trading
this today should assume the COVID row, not the 1973 row. The 20-year
production study, whose three crises include COVID, sees the same thing much
more sharply in aggregate (−12.1% → −22.3% across four weeks of lag).

The return parity is a lag-0 artifact in both studies and must never be quoted
as an edge.

Note the placebo rows: once the switch is timed honestly, simply holding the
bond sleeve every week does about as well as switching on the rate regime, and
even a pure T-bill sleeve gets most of the Sharpe. Almost all of the switch's
apparent value was the timing. The same class of error was found and fixed
independently in the production sleeve (PR #554) — that one stamped the
*defensive leg* a week late rather than the switch a week early. Note the two
studies are on different bases: this one is total return throughout, while the
production 20-year engine reads unadjusted closes (price-only), so only the
*deltas* are comparable between them, never the levels.

**2. Five more episodes.** The published study lists eight crises. The chart
adds LTCM/Asia 1998, EU debt 2011, China/oil 2015–16, Q4 2018 and the 2025
tariff shock, flagged `in_published_study: false` and rendered distinctly. They
are shown whatever they say — the two where the gate helps least (EU debt 2011,
Tariff shock 2025) are new additions, and both are on the chart.

**3. The calm column, in the same view.** Crisis windows are picked after the
fact *because* they were selloffs, so a de-risking overlay is expected to win
them; on its own that panel is a cherry-pick by construction. So the endpoint
also splits every week of the 54 years into crisis / calm and reports both:

| 1971→2026, both signals lagged, T-bill cash | weeks | buy & hold | + credit gate |
|---|---:|---|---|
| the 13 crisis windows | 661 (23%) | −16.9%/yr, −92.3% maxDD | +1.3%/yr, −25.5% maxDD |

| everything else | 2184 (77%) | **+21.5%/yr**, −16.1% maxDD | +13.9%/yr, −13.3% maxDD |
| all weeks | 2845 | 11.2%/yr, 0.72 Sharpe | 10.9%/yr, 1.22 Sharpe |

That middle row is the price of the protection: **the gate gives up ~7.5pp a
year in the calm 77% of history** (average exposure 0.54, vs 0.25 inside the
crises). The crisis windows give back almost exactly that much, which is why the
full-cycle return lands within 0.4pp of buy & hold while the drawdown halves.
Per episode the gate cuts the drawdown in **13 of 13** crises and beats the
market's return in **12 of 13** — the crisis panel was never the part that was
wrong; the full-cycle *return* claim was.

## Reproduce

```bash
cd backend
python -m scripts.long_history_credit_gate   # Fama-French + keyless FRED + IEF, cached
```

[`scripts/long_history_credit_gate.py`](../backend/scripts/long_history_credit_gate.py)
is self-contained (no DB dependency) and caches all inputs to a gitignored
scratch dir.

## Bottom line

The credit gate is not a 20-year artifact — but it is not free money either.
Over **54 years** — across stagflation, Volcker, 1987, 1990, dot-com, the GFC,
COVID and the 2022 rate bear — the same US-`NFCICREDIT` overlay roughly **halved
the drawdown** of the US market and improved risk-adjusted return in **every
decade**, while **ending with less money than buy & hold** ($226 vs $336). The
"doubled the money" version of this sentence stood here for three months and was
an artifact of three timing and sleeve errors.
Combined with the cross-market study, the gate clears both axes a skeptic asks
for — but on the *drawdown* axis only. Across markets and across time it halves
drawdowns; it does not reliably add return in either direction, and the
return-positive versions of both studies were timing artifacts.
connecting…