Skip to content

Pairs Trading

Statistical arbitrage with grouped instruments.

Concept

Pairs trading bets on mean reversion within groups:

  1. Group related stocks (sector, industry, pairs)
  2. Compute relative value (spread)
  3. Trade the spread back to equilibrium

Basic Pairs Signal

Long laggards, short leaders within each sector:

group_demean(-returns(5), 'sector')

group_demean subtracts the group mean, so: - Underperformers (negative deviation) → positive signal - Outperformers (positive deviation) → negative signal

Z-Score Normalized

Normalize the spread by its own volatility over time:

let spread = group_demean(returns(5), 'sector'),
    spread_vol = ts_std(spread, 60)
in -spread / spread_vol

group_std(returns(5), 'sector') is a different number: how far apart the sector's members are on one day. Dividing by it, group_demean(x, 'sector') / group_std(x, 'sector'), is a z-score within the sector that day, not a measure of how stretched a spread is against its own history.

Threshold Trading

Only trade when spread is stretched:

let spread = group_demean(returns(5), 'sector'),
    spread_vol = ts_std(spread, 60),
    z = spread / spread_vol
in where(abs(z) > 1.5, -spread, None)

Volatility-Matched Pairs

For pairs with different volatilities, compare risk-adjusted moves, so the quieter stock's move counts as much as the noisier one's:

group_demean(returns(5) / volatility(60), 'pair')

Partner Returns

For a group of two, the other member's return is the group sum minus your own. That gives each stock its partner's return with no extra data:

let r = returns(1),
    partner = group_sum(r, 'pair') - r
in ts_beta(r, partner, 60)

Multi-Leg Pairs

For groups with >2 members:

# Long the bottom 40% of each sector by recent return, short the top 40%

let ret = returns(5),
    grp_rank = group_rank(ret, 'sector')
in where(grp_rank < 0.4, 1, where(grp_rank > 0.6, -1, None))

Correlation Filter

Only trade correlated pairs:

let r = returns(1),
    partner = group_sum(r, 'pair') - r,
    corr = ts_corr(r, partner, 60),
    spread = group_demean(returns(5), 'pair')
in where(corr > 0.7, -spread, None)

Pairs Found from the Data

Everything above uses pairs given in the data, a 'pair' label. ts_match finds them instead: each stock's statistical twin, chosen from history, point in time (Statistical Twins). With matching='exclusive' the twins are mutual pairs, and the reversal above becomes:

let key  = delay(returns(1), 5),
    twin = ts_match(returns(5), key, 252, within='sector', matching='exclusive')
in -(returns(5) - twin)

The pairs are re-chosen every month from the year before, so they follow how stocks actually trade together, at the cost of a pair that can change at a rebalance.