Backtesting evidence

Out-of-sample: does the credit-gate overlay hold outside US sectors? (Jun 2026, corrected Sep 2026)

Versioned source evidence · docs/OUT_OF_SAMPLE_CREDIT_GATE.md

SHA-256: 46f9a12e9cbaece4d03e16893eea3d27e285a2fab0d299e690a7b45bb6e643f1

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 A superseded report retains its original assertions; it does not replace a corrected finding.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

# Out-of-sample: does the credit-gate overlay hold outside US sectors? (Jun 2026, corrected Sep 2026)

> **⚠️ Corrected Sep 2026 — same two errors found in the US 20-year study (PR #554)
> and the 54-year study (PR #558) were present here.** Both signals were stamped on
> the payoff week rather than lagged, and the cash leg booked 0.0% instead of a
> T-bill. Correcting both costs the overlay **~2.5pp of CAGR and ~0.3 of Sharpe**.
>
> **The drawdown result is essentially unchanged and still the headline.** What
> shrank is the return lift — which is exactly the pattern in the other two
> studies. The claim that the EU-sector overlay "almost exactly reproduces" the US
> result is now **false** and has been removed.

The credit-gated winner [`stovall_credit_bonds`](ROTATION_CREDIT_RESEARCH.md)
cleared every robustness test *within* its universe (US SPDR sectors, 20y) — but
the one caveat left standing was: **it's one universe and one sample; other
markets/regions are untested.** This closes that gap.

## TL;DR

**The overlay holds out of sample — as a drawdown control.** Applying the *exact*
production credit gate (equity exposure = `1 − trailing-percentile(NFCICREDIT)`,
de-risked fraction → IEF when the 10y is falling, else the T-bill) to genuinely
non-US equity bases cuts crisis drawdowns by two-thirds to three-quarters and
lifts Sharpe by ~0.4–0.5, in both universes and in every walk-forward half.

| Universe (span) | base Sharpe / maxDD | **+ overlay, corrected** | ΔSharpe | *(as published)* |
|---|---|---|---|---|
| Developed-market **country** rotation (25.8y) | 0.49 / −58.7% | **0.89 / −20.7%** | +0.40 | *1.13 / −19.7%* |
| European **sector** rotation, STOXX-600 (18.8y) | 0.43 / −50.0% | **0.90 / −15.9%** | +0.47 | *1.22 / −15.0%* |
| *(reference: US SPDR sectors, 19.9y, total return)* | 0.69 / −58.2% | *1.24 / −12.1%* | +0.55 | *1.22 / −12.9%* |

Two things to read off that table:

- **The drawdown numbers barely moved** (−19.7 → −20.7, −15.0 → −15.9). The
  alignment leak inflated *return*, not risk reduction — the same split the
  54-year study found under a publication-lag sweep, where drawdown protection
  survived and return parity did not.
- **The "EU reproduces the US" claim is gone.** Corrected, the EU overlay reaches
  Sharpe 0.90 against the US strategy's 1.24. The gate transfers, but it works
  *less* well abroad than at home, not equally well.

**Thesis still supported, with a narrower claim:** US `NFCICREDIT` is a *global*
risk-off signal — a US-credit gate de-risks foreign equity in the same crises
(2008/2020/2022) it de-risks US sectors, and the effect is not a US-sector
artifact. What it does *not* show is that the gate is an equally strong
risk-adjusted improver outside the US.

## What is (and isn't) being tested

The validated US strategy is **Stovall sector selection × credit overlay**. The
selection half (Stovall market-stage → leader sectors) is US-business-cycle
specific and doesn't port to foreign sectors or to countries. But the research's
central finding was *"the durable edge is the overlay, not the selection"* — so
the right out-of-sample test is whether **the overlay** transfers. We therefore
hold the selection deliberately generic and vary it, to prove the improvement
comes from the gate, not the picks:

- **Bases (the "WHICH"):** an equal-weight weekly-rebalanced basket (≈ passive,
  minimal selection skill) and a top-3 cross-sectional momentum rotation.
- **Overlay (the "HOW MUCH"):** the **exact** production formula — imported
  `_credit_exposure_series` and `CREDIT_RATE_MA_WEEKS` from
  [`rotation_strategies.py`](../backend/app/scoring/rotation_strategies.py), with
  the IEF/cash rate-regime routing replicated verbatim.

The overlay improves **both** bases in **both** universes, and most of the lift
shows up on the *equal-weight* base — i.e. it is the gate doing the work, not the
stock/sector picking. Numbers below are **net of 10 bps/side**.

## Results

### Developed-market country rotation (18 iShares MSCI single-country ETFs, 2001→2026)

| leg | CAGR | Sharpe | maxDD | Calmar | alpha vs EW |
|---|---:|---:|---:|---:|---:|
| EW basket (benchmark) | 8.0% | 0.49 | −58.7% | 0.14 | — |
| **EW + credit overlay** | **8.5%** | **0.89** | **−20.7%** | **0.41** | **+95%** |
| Momentum top-3 (base) | 4.8% | 0.33 | −57.7% | 0.08 | −399% |
| Momentum top-3 + overlay | 6.1% | 0.57 | −24.2% | 0.25 | −265% |

*(as published, both signals unlagged + 0% cash: EW + overlay 11.0% / 1.13 /
−19.7%; momentum + overlay 8.6% / 0.76 / −23.6%.)*

### European sector rotation (16 iShares STOXX-600 sector ETFs, Xetra, 2008→2026)

| leg | CAGR | Sharpe | maxDD | Calmar | alpha vs EW |
|---|---:|---:|---:|---:|---:|
| EW basket (benchmark) | 6.5% | 0.43 | −50.0% | 0.13 | — |
| **EW + credit overlay** | **7.9%** | **0.90** | **−15.9%** | **0.49** | **+89%** |
| Momentum top-3 (base) | 3.3% | 0.27 | −51.2% | 0.07 | −139% |
| Momentum top-3 + overlay | 6.8% | 0.70 | −19.2% | 0.35 | +19% |

*(as published: EW + overlay 10.7% / 1.22 / −15.0%.)*

**The overlay still beats its own base on return here, unlike in the US study** —
8.5% vs 8.0% and 7.9% vs 6.5%. That is not evidence it works better abroad: these
non-US bases are simply weak (Sharpe 0.49 and 0.43, −50 to −59% drawdowns), so a
sleeve that sits in Treasuries and bills through their crashes clears a low bar.
Against a strong benchmark — SPY total return, 11.08% / 0.68 — the same overlay
does *not* out-return the index. Compare gate-vs-its-own-base, never
gate-vs-someone-else's-benchmark.

### Crisis peak-to-trough (EW base → EW + overlay)

| crisis | country base | country + overlay | EU base | EU + overlay |
|---|---|---|---|---|
| GFC 2008 | −48.4% / −58.7% | **+14.0% / −7.6%** | −45.0% / −50.0% | **+6.9% / −7.3%** |
| COVID 2020 | −21.9% / −34.5% | **+10.1% / −2.5%** | −21.5% / −34.7% | **+11.5% / −2.5%** |
| Rate-bear 2022 | −20.1% / −25.7% | **−3.0% / −4.3%** | −13.4% / −19.4% | **−3.0% / −4.8%** |

The gate de-risks exactly when it matters: both bases were *positive* through the
GFC and COVID crashes once gated, and the 2022 rate-bear loss shrank to low
single digits — the same crisis behaviour seen in the US study, and the
rate-regime switch (cash, not bonds, in 2022) is what prevents the IEF sleeve
from compounding the rate-hike damage.

### Walk-forward (disjoint halves, EW base → EW + overlay)

| | country base → overlay | EU base → overlay |
|---|---|---|
| 1st half Sharpe (maxDD) | 0.48 → **0.89** (−59% → −21%) | 0.32 → **0.81** (−50% → −16%) |
| 2nd half Sharpe (maxDD) | 0.52 → **0.91** (−36% → −14%) | 0.56 → **1.01** (−35% → −9%) |

Wins risk-adjusted in **all four** disjoint sub-periods, with maxDD roughly a
third of the ungated base throughout. No window trap.

## Caveats

- **Overlay, not the full strategy.** This validates the *transferable* half (the
  credit gate). The Stovall *selection* is US-specific and was deliberately not
  ported; the foreign bases here are generic (EW / momentum), so foreign
  *absolute* returns aren't directly comparable to the US winner — the
  *improvement from the gate* is the apples-to-apples result.
- **The haven is still US Treasuries.** The de-risked sleeve routes to **IEF**
  and the rate-regime switch keys off **DGS10** — i.e. a non-US rotation gated
  this way still parks risk-off capital in US duration. That's intentional
  (flight-to-quality is global and lands in US Treasuries), but it means the
  overlay is US-macro-driven even on foreign equity.
- **Correlated samples.** 2008/2020 are global events, so US-sector, EU-sector
  and global-country results aren't fully independent draws. The consistency
  across three distinct equity universes is still strong evidence the gate isn't
  curve-fit to US sectors.
- **Signal alignment (corrected Sep 2026).** `base[w]` is the return over
  `(w-1, w]`, so both the credit exposure and the bonds-or-bills switch are now
  `.shift(1)`-ed — decided with week-`w-1` information. Unlagged, `falling@w`
  (a *backward*-looking condition: the 10y under its 26-week mean means bonds have
  already been rallying) routed the sleeve straight into the rally it was
  measuring. Worth ~2.5pp of CAGR here, matching PR #554 and PR #558.
- **The cash leg earns the 13-week T-bill (`DTB3`), not 0.0%.** While rates rise
  the de-risked fraction sits in cash, which is a bill position rather than
  banknotes; booking zero was a one-sided penalty the EW benchmark never paid.
  The two errors point in opposite directions, which is why fixing only the leak
  would leave a number wrong the other way.
- **NFCI publication lag is still unmodelled.** `NFCICREDIT` is weekly with a
  release lag; this study forward-fills onto the W-FRI grid exactly as
  `_credit_scaled_panels` does, so it shares (does not add to) the live strategy's
  lag assumption. The 54-year study swept it and found ~−0.45pp of CAGR per extra
  week of delay with drawdown protection almost untouched — so a realistic lag
  erodes the *return* result here further while leaving the drawdown result
  intact. Not swept on these universes.
- **ETF universes are non-survivorship-biased by construction** (continuously
  listed index ETFs), as with the US SPDRs.

## Reproduce

```bash
cd backend
python -m scripts.out_of_sample_credit_gate   # fetches yfinance + keyless FRED, cached
```

[`scripts/out_of_sample_credit_gate.py`](../backend/scripts/out_of_sample_credit_gate.py)
is self-contained (no DB dependency) and caches prices/FRED to a gitignored
scratch dir, so re-runs are instant and offline.

## Bottom line

The credit gate is **not** a US phenomenon, and that is the claim this study
still supports. The same US-`NFCICREDIT` overlay turns a −59%-drawdown global
country basket into a −21%-drawdown one, and a −50%-drawdown European sector
basket into −16%, winning risk-adjusted in all four disjoint halves and every
crisis window.

What it does **not** support, once the alignment and cash-leg errors are removed,
is the stronger claim it used to make: that the overlay reproduces the US result
abroad (EU 0.90 vs US 1.24), or that its value is a return lift. Across all three
studies now — US 20-year, non-US 25.8/18.8-year, and global 54-year — the
consistent, survivable finding is the same one: **the credit gate is drawdown
control that transfers across markets and decades; the return improvement was
mostly the leak.**

The residual caveat in [`ROTATION_CREDIT_RESEARCH.md`](ROTATION_CREDIT_RESEARCH.md)
is resolved in the narrow sense — the overlay does generalise — but it should be
cited for the drawdown result, not the Sharpe-1.2 headline it used to carry.
connecting…