Tickfloor

Pairs Trading: does it work?

Last updated

In-sample, not a promise. Every figure we measured here comes from a backtest on past data.

Did Pairs Trading beat the benchmark in Tickfloor's test?

No. We tested Pairs Trading on two US stock universes (171 stocks and 40 stocks), 15 ETFs, 18 crypto pairs, 2005 to 2025. After costs, it returned 2.2 to 89.0 percentage points a year less than the benchmark. None held up once we allowed for the many other rules we tested. This was a historical test, not advice.

Pairs trading picks two assets that usually move together and bets on the gap closing when it widens. Statistical arbitrage is the same idea run across many assets at once.

The rule we tested. Four versions: long-leg-only divergence on US stocks, fixed same-industry pairs reverting from a 60‑day z-score, re-convergence after a correlation breakdown on ETF pairs, and PCA-based statistical arbitrage across crypto majors.

Receipt: parameters, data, window and result file

Pairs-trading divergence, long leg only (wide US stocks)

Run id
r27-wide-pairs-lo, run 26 September 2026
Frozen parameters
universe US_STOCKS_WIDE; rebalance monthly; timeframe 1d; lag=2, formation=252, need=254
Data source
vendor not recorded in the result file; data fingerprint sha256 cc9ad92b16269979...
Instruments and coverage
171 assets: AAPL, NVDA, AVGO, ORCL, CRM, AMD, INTC, ADBE, QCOM, TXN, AMAT, PANW and 159 more (full list in the result file). 32 corrupt vendor bar-day(s) across 32 assets were dropped, never repaired
Date window
15 September 2016 to 14 March 2025 (discovery cutoff 15 March 2025)
Sample
91 holdings changes across 102 rebalances (a holdings change is a rebalance that altered what the portfolio held, not a completed trade) over 15 September 2016 to 14 March 2025 (8.5 years)
Cost rule
0.05% per side; charged on every change in position/weight; benchmark pays none; no funding/borrow on shorts
Fill rule
not recorded
Benchmark
equal-weight long of the same assets, rebalanced daily across assets trading that day (no costs)
Run record (public)
research/hunt2/ledger.jsonl, line with id r27-wide-pairs-lo
Full result file
not published yet

Fixed same-industry pairs, 60-day z-score reversion

Run id
s29-us-industry-pairs, run 29 September 2026
Frozen parameters
rebalance daily; timeframe 1d; window=60, entryZ=2, exitZ=0.5, maxHold=20, legWeight=0.0625
Data source
vendor not recorded in the result file; data fingerprint sha256 7b5251fa65d8de0e...
Instruments and coverage
40 assets: AAPL, NVDA, AMZN, GOOGL, META, TSLA, JPM, V, UNH, XOM, JNJ, COST and 28 more (full list in the result file). 9 corrupt vendor bar-day(s) across 9 assets were dropped, never repaired
Date window
15 September 2016 to 14 March 2025 (discovery cutoff 15 March 2025)
Sample
567 holdings changes across 1,933 rebalances (a holdings change is a rebalance that altered what the portfolio held, not a completed trade) over 15 September 2016 to 14 March 2025 (8.5 years)
Cost rule
0.05% per side; charged on every change in position/weight; benchmark pays none; no funding/borrow on shorts
Fill rule
not recorded
Benchmark
equal-weight long of the same assets, rebalanced daily across assets trading that day (no costs)
Run record (public)
research/hunt2/ledger.jsonl, line with id s29-us-industry-pairs
Full result file
not published yet

Correlation-breakdown re-convergence on fixed ETF pairs

Run id
corrbk-decoupling-reconvergence-etf, run 21 September 2026
Frozen parameters
rebalance daily; timeframe 1d; baseWindow=252, shortWindow=20, minBaseCorr=0.7, corrDrop=0.4, minGap=0.03, holdDays=10, maxOpen=4, legWeight=0.125
Data source
vendor not recorded in the result file; data fingerprint sha256 e417b0d42f9bb039...
Instruments and coverage
15 assets: GLD, SLV, TLT, IEF, HYG, LQD, AGG, XLK, QQQ, XLE, DBC, XLP and 3 more (full list in the result file). 2 corrupt vendor bar-day(s) across 2 assets were dropped, never repaired
Date window
4 January 2005 to 14 March 2025 (discovery cutoff 15 March 2025)
Sample
10 holdings changes across 50 rebalances (a holdings change is a rebalance that altered what the portfolio held, not a completed trade) over 4 January 2005 to 14 March 2025 (20.2 years)
Cost rule
0.05% per side; charged on every change in position/weight; benchmark pays none; no funding/borrow on shorts
Fill rule
not recorded
Benchmark
equal-weight long of the same assets, rebalanced daily across assets trading that day (no costs)
Run record (public)
research/hunt2/ledger.jsonl, line with id corrbk-decoupling-reconvergence-etf
Full result file
not published yet

PCA-based statistical arbitrage across crypto majors

Run id
mlq-pca-stat-arb-crypto, run 15 September 2026
Frozen parameters
universe CRYPTO_DAILY; rebalance daily; timeframe 1d; pcaWindow=60, zscoreWindow=20, zLongThresh=-2, zShortThresh=2, nComponents=3
Data source
vendor not recorded in the result file; data fingerprint sha256 e7098c5c15a8401c...
Instruments and coverage
18 assets: BTCUSDT, BNBUSDT, XRPUSDT, ADAUSDT, DOGEUSDT, LTCUSDT, LINKUSDT, TRXUSDT, BCHUSDT, ATOMUSDT, ETCUSDT, DASHUSDT and 6 more (full list in the result file)
Date window
18 August 2017 to 14 March 2025 (discovery cutoff 15 March 2025)
Sample
735 holdings changes across 429 rebalances (a holdings change is a rebalance that altered what the portfolio held, not a completed trade) over 18 August 2017 to 14 March 2025 (7.6 years)
Cost rule
0.10% to 0.17% per side; charged on every change in position/weight; benchmark pays none; no funding/borrow on shorts
Fill rule
not recorded
Benchmark
equal-weight long of the same assets, rebalanced daily across assets trading that day (no costs)
Run record (public)
research/hunt2/ledger.jsonl, line with id mlq-pca-stat-arb-crypto
Full result file
not published yet

No. None of the versions we tested passed. None held up once we allowed for how many other rules we tested. That covers the versions, costs and benchmark described here, not every version of the rule.

Tested a rule of your own? Check whether your backtest is just luck. It is free and needs no account.

In the stored result for Correlation-breakdown re-convergence on fixed ETF pairs, the rule's return is exactly 0.0% from 4 January 2005 to 5 February 2015, the first half of the window. Tickfloor has not established why, so read the full-period figure as a weaker test than its length suggests.

This is one line in a bigger check. The research desk has run 1,346 strategies, most against an equal-weight benchmark of the same assets rebalanced daily that pays no costs (the strategies pay theirs). After correcting for how many we tried (the luck check, Benjamini‑Hochberg, 1,778 tests), 0 passed. That does not prove no strategy works, and it says nothing about versions we did not test.

Short answers: the verdict and the rule, as questions

Does pairs trading work?

No. None of the versions we tested passed. After costs: US stocks −2.2 pts/yr vs the benchmark; US stocks −19.5 pts/yr vs the benchmark; ETFs −7.1 pts/yr vs the benchmark; crypto pairs −89.0 pts/yr vs the benchmark. None held up once we allowed for how many other rules we tested. That covers the versions, costs and benchmark described here, not every version of the rule.

What exact rule did Tickfloor test?

Four versions: long-leg-only divergence on US stocks, fixed same-industry pairs reverting from a 60-day z-score, re-convergence after a correlation breakdown on ETF pairs, and PCA-based statistical arbitrage across crypto majors.

Is this financial advice?

No. This measures a publicly claimed strategy rule, not a recommendation. General information only, not personal advice.

See the full numbers on the Mean reversion, Machine learning family pages, or the full method and every result.

Other rules we've tested

Terms explained

Next steps

General information only. It doesn't consider your objectives, finances or needs. Tickfloor holds no financial services licence and never places trades.