How to use the Pair Trading Strategy agent skill

What the pair-trading skill makes the agent do

An agent skill is a folder containing a SKILL.md instruction file and any supporting files. When a skill matches your request, the agent reads its instructions and follows them. This one covers pair trading: choosing two highly correlated instruments, measuring how far their price ratio deviates from the mean, and trading against extreme deviations while waiting for the ratio to revert.

The signal logic is simple. The agent computes the ratio of the two closing prices, takes a rolling mean and standard deviation over a lookback window, and converts the ratio into a Z-score. A Z-score below the negative entry threshold means going long A and short B; above the positive threshold means short A and long B; and a Z-score back inside the exit threshold closes the position.

• Ratio: close_A / close_B
• Z-score: (ratio - rolling mean) / rolling standard deviation
• Default parameters: lookback 60, entry_z 2.0, exit_z 0.5
• Sizing: equal weight, with A and B each taking 50% of capital

Install the pair-trading skill

You can install the skill in one of three ways. Download the ZIP from the skill's page on this site and unzip it into your personal skills folder at ~/.claude/skills/, or into a project's .claude/skills/ folder if you want it available only in that project. Alternatively, upload the skill in the Claude apps under Settings → Capabilities → Skills.

Make sure the skill folder is named pair-trading and contains SKILL.md, since the skill is identified by its name in that file.

• Personal install: ~/.claude/skills/pair-trading/
• Project install: .claude/skills/pair-trading/ inside your project
• Claude apps: Settings → Capabilities → Skills, then upload

Set up the requirements

The skill needs Python with pandas and numpy. Install them with the command below. You need a market data source such as Tushare (for stocks) or OKX (for crypto) only if you run the example configuration; the logic itself does not depend on a particular data provider.

• Python with pandas and numpy
• A market data source such as Tushare or OKX, only for the example config

Trigger the skill with natural requests

You do not need special syntax. Describe the pair and what you want, and mention pair trading or mean reversion so the agent matches the skill. Name exactly two instruments, because the skill requires at least two and its code expects exactly two.

Good requests say what you want produced: a signal script, a backtest, or an explanation of the Z-score logic.

• "Use the pair-trading skill to build signals for 601318.SH and 601628.SH."
• "Draft a mean-reversion pair strategy on BTC-USDT and ETH-USDT with a 60-day lookback."
• "Explain how the Z-score entry and exit thresholds are calculated for a pair of stocks."

Configure the example run

The skill ships with an example config.json. The stock example uses the tushare source with two Chinese stock codes and a date range from 2023 to 2024. The crypto version uses the okx source with BTC-USDT and ETH-USDT. Both set initial_cash to 1000000 and commission to 0.001, and both have extra_fields set to null.

Check the codes array first. It must contain exactly two instruments, and the date indexes of the two must be aligned with an inner join before the ratio is calculated.

• codes: exactly 2 instruments
• Dates: the two series must align (inner join)
• Early rows have NaN Z-scores until the lookback window fills, so signals are filled with 0

Read the signal convention

The skill reports positions as signed weights. Instrument A is 0.5 when long, -0.5 when short, and 0 when flat. Instrument B always moves in the opposite direction to A. The skill also warns against generating same-direction signals for both legs, because pair trading is a long-short hedge.

• A: 0.5 long, -0.5 short, 0 flat
• B: opposite direction to A

Safety notes: scripts and network

The skill runs scripts: its package includes example_signal_engine.py, a short Python reference implementation. Read the script before running it, especially if you run it on your own machine. The skill does not use the network itself; the market data source is only needed if you run the example config.

The skill has no risk controls, no slippage modelling and no stop-losses. Its logic uses fixed parameters and equal-weight sizing, and it can lose money when the correlation between the two instruments breaks down. Do not use it for live trading with real capital without independent validation.

• Review example_signal_engine.py before running it
• No built-in stop-losses, risk limits or slippage model
• Not validated for live capital without your own testing

Tips for better results

Choose pairs that are genuinely correlated, such as two stocks in the same industry or BTC and ETH, because the logic assumes the ratio tends to revert. Test different lookback, entry_z and exit_z values in your own backtest rather than relying on the defaults. Keep the two legs aligned by date before you compare them.

• Pick pairs with a economic or industry link, not random tickers
• Vary lookback, entry_z and exit_z in your own tests
• Verify date alignment and that codes has exactly two entries