FX cointegration pairs trading

Replication code and Typst manuscript for Lemishko, Landi & Caicedo-Llano, Cointegration-Based Strategies in Forex Pairs Trading (August 2026 rewrite; frozen sample through 2025-12-31). Working-paper landing page: SSRN. If the SSRN PDF lags this repository, treat paper/main.typ / paper/main.pdf and the frozen parquet here as authoritative for current numbers.

Please cite this repository via GitHub Cite this repository / CITATION.cff.

Setup

Requires uv and Python ≥ 3.12.

git clone https://github.com/QuantLandi/fx-cointegration.git
cd fx-cointegration
uv sync

Frozen prices are in data/fx_prices.parquet. Re-download only if you need a new freeze:

uv run python scripts/01_download_prices.py

Pipeline

From the repo root:

uv run python scripts/02_backtest.py --clear-panels           # EG → outputs/coint/
uv run python scripts/02_backtest.py --simple --clear-panels  # → outputs/simple/
uv run python scripts/03_compare_strategies.py               # → outputs/compare/
uv run python scripts/04_portfolio_tables.py                 # → outputs/paper/tables|portfolio
uv run python scripts/05_plot_figures.py                     # → outputs/paper/figures/

04_portfolio_tables.py also refreshes pair metrics.csv from existing panels (including the {0,1,2,5} bp cost grid), so you need not re-run EG just for metrics/portfolio tables.

Runtime: full 02–05 on a laptop is typically on the order of tens of minutes (Engle–Granger screens dominate; --simple is faster). Steps 03–05 are quick once panels exist.

Data

Daily Yahoo Finance FX spots, 2007-01-01 to 2025-12-31, seven USD crosses (AUD, CAD, CHF, EUR, GBP, JPY, NZD), all quoted as XXXUSD.

Method (sketch)

1. For each undirected pair (21 = C(7,2), alphabetical legs) and rolling train/test window, screen log prices with Engle–Granger at 5%: both OLS orientations are tested; among passes, keep the clearer residual ADF (more negative t-stat) and map the hedge into log_1 − β·log_2.
2. If cointegrated, form that spread and standardize with train mean/SD for an OOS z-score.
3. Trade when |z| exceeds a threshold (long spread if z < −z*, short if z > z*). Signals are lagged two days.
4. Strategy return = signal × (r₁ − r₂) (equal notional, not β-hedged). Annualized Sharpe uses all calendar days (flat days as 0), so volatility is diluted when often out of market. Metrics trades = days with nonzero signal, not round-trips.
5. Transaction costs: round-trip κ bp of pair notional charged as (κ/1e4)·|Δsignal|/2 (open/close = κ/2 each; flip = κ). Grid {0, 1, 2, 5} bp; headline κ = 2. Panels store gross returns; costs are applied when building metrics and portfolio tables.
6. Paper portfolio: sum the 21 pair daily returns, divide by 21. Main tables are unlevered. A companion table scales each strategy ex-post to 10% annualized vol so return/MDD levels are comparable (Sharpe unchanged). Cumulative-return figures scale the EG path to equal ex-post daily vol vs simple (visuals only). Sortino uses the std of strictly negative daily returns; Calmar = ann return / |max DD|.

Paper subset: train=257, test=21, z* ∈ {1, 2, 3}, 21 undirected pairs, cost κ ∈ {0, 1, 2, 5} bp (baseline 2). Simple benchmark: --simple (no EG gate).

Outputs

outputs/
  coint/
    metrics.csv (+ .meta.json)   # rows × z × cost_bp; panels are gross
    panels/{leg1}_{leg2}/        # prices, returns, spread, zscore, signal, …
  simple/
    metrics.csv (+ .meta.json)
    panels/{leg1}_{leg2}/
  compare/
    metrics.csv                  # EG vs simple join (ΔSharpe), keyed by cost_bp
  paper/
    tables/                      # unlevered / target-vol / cost-sensitivity CSVs
    portfolio/                   # daily / cumulative series at baseline κ=2
    figures/                     # fig01–fig05

Under EG, spread / zscore are NaN outside Engle–Granger-pass OOS blocks. Under --simple, almost all OOS blocks are filled (NaN only before the first window or if train std is zero). Strategy return panels are gross (zero cost); flat days are 0. Net returns apply (κ/1e4)·|Δsignal|/2 in metrics and 04.

outputs/ and local/ are gitignored. Optional local checks (JAE table tolerances, notebook spot checks) live under local/.

Manuscript: paper/main.typ (compiled PDF: paper/main.pdf; figures via paper/figures, regenerated from outputs/paper/figures).

License

• Code, scripts, pyproject.toml, uv.lock, and data/: MIT
• Manuscript assets under paper/ (Typst source, PDF, figures, bibliography): CC BY 4.0

The frozen parquet in data/ is a derived replication sample built from Yahoo Finance FX spots; underlying Yahoo quotes remain subject to Yahoo’s terms of use.