Work
Individual ResearchPairs Trading · Walk-Forward Testing · Robustness

Statistical Arbitrage & Walk-Forward Validation

The original result did not survive a broader development test. Fresh confirmation later turned positive, but uncertainty remained too wide for a strong claim.

Central findingFresh confirmation was positive but statistically inconclusive.

Evidence

The result in context

Original total return
+20.72%
Robust-development total return
−9.41%
Frozen-confirmation return
+3.83%
Bootstrap Sharpe interval
[−0.88, 2.00]

Question

Does a promising pairs-trading result persist across more pairs, groups, market periods and genuinely out-of-sample observations?

A pairs strategy taken from an encouraging small baseline through broad walk-forward development and a frozen confirmation period.

Original strategy

A small result that needed a larger test

The baseline returned 20.72% with a 0.68 Sharpe ratio across 29 trades, five unique pairs and three groups. That breadth was too narrow to treat the result as stable evidence.

Robust development

Walk-forward testing reversed the headline

Across 2015–2024, broader out-of-sample development covered 199 trades, 62 pairs and 17 groups. Total return was −9.41% with a −0.30 Sharpe ratio.

The reversal is informative: a strategy that works in a small selected setting may not survive wider opportunity sets and repeated re-estimation.

Research evolution

From small baseline to fresh evidence

Each stage increases breadth or separation from development decisions.

Total return
+20.72%
Sharpe ratio
0.68
Maximum drawdown
−9.06%
Trades
29
Breadth
5 pairs · 3 groups

Encouraging, but too narrow to establish robustness.

Interactive evidence

The full development study reversed the archived headline

Cumulative growth of £1 from 1,886 committed daily strategy returns.

Read: Development finishes at 0.906×, a −9.41% total return, after a maximum drawdown of roughly 11.12%.

Data table · 1,886 verified rows
Representative preview rows; the full 1,886-row local dataset loads with the interactive chart.
DateStrategy ReturnEquityDrawdown Pct
2017-07-05010
2017-07-27-0.0076190.980827-2.071819
2018-04-300.0009530.954656-4.684865
2019-02-250.0011440.95475-4.675439
2020-01-10-0.0016250.951232-5.026683
2020-11-03-0.0004430.966195-3.532781
2021-08-270.0004020.941605-5.987891
2022-06-230.0013890.927207-7.425437
2023-05-110.0002320.930491-7.097552
2024-03-070.0059750.909941-9.149299
2024-12-060.0006620.909709-9.17248
2024-12-31-0.0004850.90586-9.556731

Frozen confirmation

The sign improved; the confidence did not

From 2025 to July 2026, return was +3.83% with a 0.59 Sharpe ratio across 43 trades, 19 pairs and 12 groups. One pair produced 28.34% of absolute confirmation P&L.

The bootstrap Sharpe interval spans −0.88 to 2.00 and the Newey–West p-value is 0.421. Fresh confirmation was positive but statistically inconclusive.

Interactive evidence

Confirmation turned positive without resolving uncertainty

Cumulative growth of £1 from 378 committed daily returns, 2 Jan 2025–8 Jul 2026.

Read: The frozen confirmation window ends at 1.038× (+3.83%); the evidence remains too short for a strong performance claim.

Data table · 378 verified rows
Representative preview rows; the full 378-row local dataset loads with the interactive chart.
DateStrategy ReturnEquityDrawdown Pct
2025-01-02010
2025-01-08010
2025-03-100.0008820.994571-1.108011
2025-05-05-0.0004510.988089-1.752524
2025-07-0900.98433-2.126352
2025-09-0400.98686-1.8748
2025-11-04-0.0030510.997027-0.863814
2026-01-020.00021.000446-0.523892
2026-03-0901.041589-0.353095
2026-05-04-0.0006841.052239-0.413169
2026-07-01-0.0006641.036317-1.920112
2026-07-080.0004131.038338-1.728852

Interactive evidence

Confirmation P&L remained concentrated in a small number of pairs

Absolute share of pair-level weighted P&L; all 19 selected pairs are retained in the data table.

Read: MS/NTRS accounts for 28.34% of absolute pair P&L, making diversification risk visible alongside the positive aggregate result.

Data table · 19 verified rows
Complete dataset
PairWeighted PnlAbsolute Pnl Share Pct
MS/NTRS0.0398328.343109
NXPI/QCOM-0.02285116.261303
COF/MA0.01850713.169421
MU/QCOM0.0122368.707019
META/MSFT0.0086396.147878
COP/OXY-0.0081195.777591
ADSK/NOW0.0062054.41556
INTU/NOW-0.0056254.002895
IEFA/VEA-0.004473.180934
COF/V-0.0026141.860104
IYW/QQQ-0.0023031.63869
C/WFC0.0022141.575463
AMZN/MSFT-0.0018491.315811
INTU/MSFT-0.0017821.268292
GS/NTRS0.0011670.830569
IYR/VNQ-0.0010270.730724
QQQ/XLK0.0009650.686493
VFH/XLF-0.0001240.088143
VDE/XLE00

Limitations

What this evidence does not establish

  • The confirmation sample is short and contains only 43 trades.
  • Pair-level contribution remains concentrated, and the statistical intervals include weak or negative outcomes.
  • The confirmation protocol and result were committed together, so the public Git history alone cannot verify the intended pre-result freeze.

Source and reproducibility

Trace the evidence

Source code, evaluation outputs and supporting material are available in the repository.

View repository
  1. Development equity and drawdownreports/development/final/portfolio_daily_returns.csvCommit / evidence ID: 7f7ac62376630cd5eb715e38fbc87ab316f1747e
  2. Confirmation pair contributionsreports/confirmation/pair_contributions.csvCommit / evidence ID: 7f7ac62376630cd5eb715e38fbc87ab316f1747e