Pairs trading, from selection to decision.

1. Start with an economic relationship

Choose assets whose relationship you can explain: companies exposed to similar activities, or funds following comparable holdings. Check liquidity, trading hours, currency and the feasibility of borrowing the short leg. A shared sector is a reason to investigate, not proof of an equilibrium.

Pairs trading is a family of strategies. The distance method compares normalized price paths. A cointegration approach tests a weighted relationship between prices. PairScanner uses daily-return correlation to shortlist candidates before testing cointegration.

2. Separate correlation from cointegration

Correlation describes how returns move together. Two assets can have correlated returns while their price relationship drifts. PairScanner tests log prices with the Engle–Granger method and estimates a hedge coefficient, β.

The null hypothesis is no cointegration. A small p-value is evidence against that hypothesis under the test’s assumptions; it is not the probability that a pair will fail. Testing many pairs creates false discoveries, which is why the scanner also exposes a q-value adjustment and a chronological validation check. Engle–Granger test documentation.

3. Read the spread and the Z-score

Spread = log(A) − β × log(B)
Z = (spread − rolling mean) / rolling standard deviation

PairScanner normally uses 60 sessions for the rolling Z-score. For positive β, a negative Z suggests A is relatively low against B under the fitted model; the reverse applies to a positive Z. A value around ±2 is a common research threshold, not an instruction to trade.

The coefficient estimated on log prices describes relative sensitivity. It is not a share count, and a one-dollar-versus-one-dollar allocation does not automatically reproduce that hedge. Check orientation, sizing, market exposure and model stability before interpreting a signal.

4. Check a later period

An attractive fit on the same data used to select the pair can be misleading. In the published case studies, the chronological check fits β on the earlier 70% of observations, freezes it and examines the remaining period. The individual study states its exact dates, tests and limitations.

Passing a statistical check is different from demonstrating a profitable strategy. A trading backtest must also specify executable entry and exit times, transaction costs, financing, short availability and position sizing. Choosing an example after seeing its result introduces selection bias even when a later-period check is reported.

5. Compare three different conclusions

These studies were frozen in September 2026. KO/PEP did not pass the initial cointegration threshold. MBB/SPMB passed the reported checks with a Z-score near zero. KMI/WMB combined a passed check with a larger divergence. These are historical readings, not current signals.

6. Add context and define the exit

Read issuer filings and investigate news that could explain the divergence. A rise in article counts only tells you that coverage changed; it does not establish why prices moved. Check whether the event occurred before or after the data timestamp.

Write down an exit condition, a maximum holding period and a loss limit before any position. Review exposure across all pairs: repeatedly buying the same stock in several pairs can concentrate risk. Keep a record of rejected ideas as well as accepted ones.

Use the metric definitions for the exact vocabulary, or explore the scanner workflow.