← Lab
S034

Pairs / stat-arb on cointegrated ETF spreads

pairs-cointegration-etf-v1
failed Stage 1 (Quick-screen) 2026-07-03
Share: Twitter / X LinkedIn

ID: S034 Slug: pairs-cointegration-etf-v1 Run date: 2026-07-03 Failed at: Stage 1 (Quick-screen) Outcome: FAILED — all 7 cells NULL. Headline metric: cointegration p-values 0.06–0.99 (gate needs < 0.05); best cell Sharpe 0.15 (gate ≥ 0.8) Fail reason: None of the seven economically-tethered ETF pairs is cointegrated over 2010–2026 — not with the fixed β=1 log-ratio, and not with a fitted hedge ratio either. The spreads trend rather than mean-revert, so the cointegration gate correctly refuses to trade almost every window; what little does trade earns no risk-adjusted edge. The discipline works exactly as designed — it just finds nothing here.

What we found

Cell Pair ADF β=1 EG fitted-β Cointegrated folds Trades Sharpe β-SPY Max-DD Verdict
c01 GLD/SLV 0.33 0.74 4% 3 −0.23 0.00 4.9% NULL
c02 IEF/TLT 0.50 0.99 5% 4 −0.03 0.00 4.3% NULL
c03 SPY/QQQ 0.91 0.24 3% 2 −0.41 0.00 5.6% NULL
c04 IWM/QQQ 0.96 0.07 5% 5 +0.09 0.00 4.1% NULL
c05 IWM/SPY 0.86 0.06 5% 5 −0.31 0.00 4.8% NULL
c06 XLE/XOP 0.83 0.98 9% 6 −0.17 0.00 11.5% NULL
c07 HYG/LQD 0.80 0.73 13% 12 +0.15 0.00 4.0% NULL

The spread-drift charts above show the mechanism directly: each "plausible" pair's price ratio wanders away from its long-run mean and stays there rather than oscillating around it — a trend, not a reversion.

Why the plausible pairs still trended (2010–2026)

Cointegration, where it appears at all, is regime-fleeting — present in one four-year window and absent in the next — which validates the per-fold re-confirmation design but is too intermittent to build a strategy on.

The methodological finding (a control worth recording)

We validated the test machinery on two ETFs that must be cointegrated — SPY vs IVV and SPY vs VOO, all S&P-500 trackers. Result: they fail the fixed-β=1 ADF (p ≈ 0.94 / 0.82) yet are overwhelmingly cointegrated with a fitted β (Engle-Granger p = 0.0000). So the frozen β=1 log-ratio construction is genuinely too restrictive — it can reject even textbook cointegration. But the fitted-β test that correctly passes SPY/IVV still rejects all seven economic pairs — so the NULL is not a β=1 artifact; these pairs are genuinely not cointegrated over this sample.

The deeper lesson is the fundamental tension of liquid-ETF stat-arb: the pairs that are truly cointegrated (SPY/IVV) have essentially no spread to trade (they are the same thing), while economically-tethered-but-distinct pairs have a tradeable spread but are, at best, fleetingly cointegrated. The tradeable-and-stationary sweet spot is absent from this preset set.

Verdict & what's next

All seven cells are NULL — cut at Stage 1 on the cointegration gate. The pillar's discipline is vindicated (it refused to trade non-reverting spreads and lost nothing), and the market-neutrality machinery is confirmed. Two things the run establishes for the next experiment: (1) the fixed β=1 construction must be replaced with a fitted-β / Johansen ratio (the SPY/IVV control proves β=1 alone can miss real cointegration); (2) pair selection matters more than the ratio — the next set should target tighter mechanical tethers where a tradeable, persistent cointegration can actually exist. Both are pre-registered as a proposed follow-up (lab/candidates/pairs-fitted-beta-johansen-v1.md).

Baseline was zero (market-neutral has no beta to lift it). Friction: 4 transactions/round-trip + modeled 0.5%/yr borrow. Data: Tiingo EOD 2010–2026. English per repo convention. Ledger: pairs-pillar-ledger.md.

What we tested — the recipe

What we tested — cointegrated ETF spreads
dollar-neutral log-ratio pairs, traded only where the spread is stationary
Universe — 7 economically-tethered ETF pairs (Tasty presets + 2)7 pairsGate — spread stationary (ADF p<0.05), re-confirmed every foldcointegrationSignal — z-score of the log-ratio, entry |z|>2, exit |z|<0.5mean-revertTail — divergence stop |z|>3.5 or a mid-trade cointegration breakstopFriction — 4 transactions/round-trip + modeled borrowdouble costGate: expectancy>0 · Sharpe≥0.8 · |beta-SPY|≤0.15 · max-DD≤15% · baseline ZERO · 2010→2026
Market-neutral by construction. The cointegration gate is the discipline — a pair that is not stationary is simply not traded.

Slice & dice

Cointegration scorecard — fraction of folds that pass the ADF gate
how often each pair's spread was stationary (higher = more tradeable); none is reliably cointegrated
GLD/SLVIEF/TLTSPY/QQQIWM/QQQIWM/SPYXLE/XOPHYG/LQDflat = 6
Even the 'plausible' pairs pass in only 3–13% of rolling windows. On the full sample every pair fails both the fixed-β=1 ADF and the fitted-β Engle-Granger test (p all > 0.05).
Spread drift — GLD/SLV price ratio vs its long-run mean
ADF β=1 p=0.33 · Engle-Granger fitted-β p=0.74 — not stationary; the ratio trends, it does not revert
2.925.147.379.5911.81201020122014201620182020202220242026GLD/SLV ratiofull-sample mean
Green = the actual price ratio; dashed = the long-run mean a mean-reverting spread would return to. Instead it drifts and stays away — the structural trend the cointegration gate correctly refuses to trade.
Spread drift — IEF/TLT price ratio vs its long-run mean
ADF β=1 p=0.50 · Engle-Granger fitted-β p=0.99 — not stationary; the ratio trends, it does not revert
0.720.830.941.051.16201020122014201620182020202220242026IEF/TLT ratiofull-sample mean
Green = the actual price ratio; dashed = the long-run mean a mean-reverting spread would return to. Instead it drifts and stays away — the structural trend the cointegration gate correctly refuses to trade.
Spread drift — XLE/XOP price ratio vs its long-run mean
ADF β=1 p=0.83 · Engle-Granger fitted-β p=0.98 — not stationary; the ratio trends, it does not revert
0.100.180.260.340.42201020122014201620182020202220242026XLE/XOP ratiofull-sample mean
Green = the actual price ratio; dashed = the long-run mean a mean-reverting spread would return to. Instead it drifts and stays away — the structural trend the cointegration gate correctly refuses to trade.
Spread drift — HYG/LQD price ratio vs its long-run mean
ADF β=1 p=0.80 · Engle-Granger fitted-β p=0.73 — not stationary; the ratio trends, it does not revert
0.540.590.640.690.75201020122014201620182020202220242026HYG/LQD ratiofull-sample mean
Green = the actual price ratio; dashed = the long-run mean a mean-reverting spread would return to. Instead it drifts and stays away — the structural trend the cointegration gate correctly refuses to trade.
← Olderpairs-cointegration-etf-v1 Newer →etf-momentum-rotation-12-1-v1

Mechaniq provides information, not investment advice. We do not execute trades. Past results are no guarantee for future performance. You are solely responsible for your trading decisions.

Lab · Methodology · Greeks Lab · About · FAQ · Glossary · Privacy · Terms · mechaniq.trade © 2026

07 Jul 2026, 07:07