Skip to content
All library documents

Choosing and Calibrating Measures of Mean Reversion

Article Quant Q&A · Author: Joshua Chance
Read the original

Summary

This exchange asks whether directional trading has a standard way to quantify mean reversion, suggesting autocorrelation, variance ratios, and the Hurst exponent as candidates. The answers offer no universally preferred measure. They mention simple time-series z-scores as one practical starting point and rolling unit-root tests as another possible approach, while emphasizing that any method must fit the problem.

Calibration depends on the security and investment horizon, including the window used to estimate behavior and the persistence of mean reversion. Rolling tests can lag, and changing market distributions raise the further question of whether parameters should adapt over time. Leverage also affects the strategy’s practical design. The discussion provides methodological guidance rather than empirical comparisons: it does not show which measure performs best, prescribe parameter values, or demonstrate that a mean-reversion signal will generate profitable trades.

Key ideas

  • There is no universally standard measure for quantifying mean reversion in directional trading.
  • Simple time-series z-scores may be a reasonable starting point when extra complexity is not justified.
  • Measure and parameter choices should reflect the specific security and investment horizon.
  • Rolling unit-root tests can lag, and their usefulness depends on the window and the persistence of reversion.
  • Non-stationary behavior may make dynamic parameters and leverage choices relevant.

Tags

Mean reversionStatisticsTechnical indicatorsRisk management
Full text
# Is there a standard method for quantifying mean-reversion for use in directional trading?


# Is there a standard method for quantifying mean-reversion for use in directional trading?












Assuming a directional strategy (no pairs or spread trades) is there a "standard" method for quantifying mean-reversion? Should auto-correlation, variance ratios, hurst exponent, or some other measure be preferred in all cases, or are there advantages to each given the context?

## Answer by ZAxisMapping (score 10, accepted)

https://quant.stackexchange.com/a/541

There is no standard method and many techniques can work well, including simple time series z-scoring. I'm many cases, I would recommend using the simpler approaches unless the added complexity can be justified.

However, the challenge with all techniques is the proper calibration, which is very much context sensitive. The parameter selection needs to be guided by the characteristics of the specific security and investment horizon. Furthermore, should the parameters themselves be dynamic? Nearly all processes are driven by non-stationary distributions. The amount of leverage used may also be an important consideration.

## Answer by RockScience (score 2)

https://quant.stackexchange.com/a/535

As usual, no standard method. You can maybe test for unit root over a rolling window, but as usual (again) you are going to be lagging. It will all depend on your choice of window and of the persistence of mean reversion in the market you consider.

Shown in full with attribution under the source's licence. Licence: CC BY-SA 4.0 (Stack Exchange)

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.