AI Pairs Trading: Build a Mean Reversion Bot

Pairs trading is a classic market-neutral strategy. Instead of betting on the direction of a single asset, you bet on the relationship between two assets. When that relationship breaks down temporarily, you trade the spread, expecting it to converge.

AI can make pairs trading more scalable by screening large universes and predicting when a spread is likely to revert.

The Idea Behind Pairs Trading

Find two assets whose prices historically move together. When their price ratio or spread deviates significantly from the mean:

• Buy the underperforming asset
• Sell the outperforming asset
• Close the trade when the spread converges

This is a market-neutral approach because profit comes from the relative movement, not the overall market direction.

Correlation vs Cointegration

Correlation measures how two assets move together over a period. Cointegration tests whether a linear combination of the two assets is stationary, meaning the spread tends to return to a long-term mean.

Cointegration is more important than correlation for pairs trading. Two assets can be highly correlated but not cointegrated, making the spread unstable.

Finding Cointegrated Pairs

Use the Augmented Dickey-Fuller test on the spread. In Python:

import statsmodels.api as sm
from statsmodels.tsa.stattools import adfuller
 
def test_cointegration(x, y):
    x = sm.add_constant(x)
    model = sm.OLS(y, x).fit()
    residuals = model.resid
    adf_result = adfuller(residuals)
    return adf_result[1]  # p-value
 
p_value = test_cointegration(prices['AAPL'], prices['MSFT'])

A low p-value, typically below 0.05, suggests cointegration.

Building the Spread

The spread is usually defined as the residuals from an OLS regression:

def calculate_spread(x, y):
    x = sm.add_constant(x)
    model = sm.OLS(y, x).fit()
    return model.resid
 
spread = calculate_spread(prices['AAPL'], prices['MSFT'])

When the spread is far from its mean, a trading opportunity may exist.

Entry and Exit Signals

A common approach:

• Enter when the z-score of the spread exceeds ±2
• Exit when the z-score returns to 0
• Stop out if the spread continues to diverge beyond ±3

spread_mean = spread.mean()
spread_std = spread.std()
z_score = (spread - spread_mean) / spread_std

How AI Improves Pair Selection

Instead of manually testing every pair, AI can:

• Screen thousands of pairs for cointegration and correlation
• Cluster assets by sector and behavior
• Detect regime changes that break historical relationships
• Predict the probability of spread convergence

A machine-learning model might use features such as:

• Cointegration p-value
• Half-life of mean reversion
• Volatility ratio
• Recent divergence magnitude
• Sector similarity

Risk Management

Pairs trading is not risk-free. Key risks include:

• Structural break: One company changes fundamentally, destroying the historical relationship
• Execution risk: Slippage on two legs can eliminate profits
• Liquidity risk: One asset may be hard to short
• Overfitting: Pairs found by mining data may not work in the future

Risk rules:

• Test pairs on multiple time periods
• Avoid pairs without an economic link
• Use position limits
• Monitor for structural breaks

Example Workflow

1. Download price data for a universe of stocks
2. Test cointegration for all pairs
3. Filter pairs with low p-values and economic rationale
4. Calculate spread and z-score
5. Backtest entry and exit rules
6. Paper trade before live deployment

Worked Example: A Pairs Trade Walkthrough

Suppose over the past year Pepsi and Coca-Cola have moved together with an ADF test p-value of 0.02. The current spread is 2.3 standard deviations above the mean. Coca-Cola has outperformed Pepsi by 4.5% over the last ten trading days.

A pairs trader might:

1. Short Coca-Cola with $5,000 notional.
2. Buy Pepsi with $5,000 notional.
3. Set a profit target when the z-score returns to 0.
4. Set a stop loss when the z-score reaches 3.5.

If the spread reverts, the trade captures the relative convergence. If a structural break occurs, such as a takeover rumor in one company, the stop loss exits the trade before the relationship completely breaks down.

When Pairs Trading Works Best

Pairs trading tends to perform well when:

• The two assets have a clear economic link, such as competitors in the same industry.
• Volatility is moderate, not spiking in one direction.
• Short availability is reliable and borrow costs are low.
• The portfolio contains many pairs so no single failure is catastrophic.

Common Pairs Trading Mistakes

• Mining too many pairs: Testing every combination in a 500-stock universe will find spurious relationships.
• Ignoring earnings calendars: A single earnings report can destroy a historical relationship overnight.
• Equal notional sizing: Pairs should be sized by volatility or beta, not by dollar amount.
• No stop loss: Mean reversion is not guaranteed. Define a point where the trade is wrong.

Bottom Line

Pairs trading is a market-neutral strategy that relies on statistical relationships rather than market direction. AI makes it possible to screen large universes and adapt to changing conditions. However, the strategy requires careful pair selection, execution discipline, and ongoing monitoring for structural breaks.

Related reading: AI Trend Following vs Mean Reversion | AI Breakout Strategy Setup | AI Strategy Comparison Framework