Half-Life: Speed of Mean Reversion

The Half-Life of mean reversion measures how quickly prices tend to return to their average level. A shorter half-life means faster reversion, which is generally more attractive for trading strategies.

The Ornstein-Uhlenbeck Process

Half-Life is calculated using the Ornstein-Uhlenbeck (OU) model, which describes mean-reverting processes mathematically. The OU process assumes prices are pulled back toward a mean level with a certain strength.

dP = θ(μ - P)dt + σdW Where: θ = speed of reversion, μ = mean level, σ = volatility

The Half-Life is derived from the speed parameter θ:

Half-Life = ln(2) / θ ≈ 0.693 / θ

Practical Interpretation

A Half-Life of 10 bars means that, on average, half of any deviation from the mean will be corrected within 10 bars. Shorter half-lives indicate stronger, faster mean reversion.

Half-Life Scoring
  • HL ≤ 50 bars - Score: 100 (fast reversion)
  • HL = 100 bars - Score: 50 (moderate)
  • HL ≥ 200 bars - Score: 0 (too slow)

Calculation Method

QuanterLab estimates Half-Life through linear regression:

  1. Calculate the spread between price and its moving average
  2. Regress the change in spread against the spread level
  3. The slope coefficient gives the reversion speed
  4. Convert to Half-Life using the formula above

Why Half-Life Matters

A security might be mean-reverting, but if the Half-Life is 500 bars, you'd have to hold positions for extended periods waiting for reversion. Shorter half-lives allow for:

  • More frequent trading opportunities
  • Reduced exposure time and risk
  • Better capital efficiency
Limitation

Half-Life is estimated from historical data and assumes the reversion process remains stable. Regime changes can alter the effective Half-Life significantly.

Platform Integration

Half-Life receives a 15% weight in the daily composite score. Securities with half-lives exceeding a maximum threshold (typically configurable based on timeframe) are filtered out as unsuitable for mean reversion strategies.

Back to QuanterLab
Report
Loading report...
Article
Loading article...