Like what you see? Follow Adrian on Twitter to be notified of new content.

I am frequently calculating the half-life of a series to use in mean reversion models. I came across this Stack Exchange answer which showed the derivation of a series half-life based on a regressible model. This is such a common calculation for me that I’ve written out the derivation here so that I never forget.

Model

Define the weakly stationary process \(X\) as: $$ \begin{align} X_t = c + \lambda X_{t-1} + \epsilon_t, \quad 0 < \lambda < 1 \label{1} \end{align} $$

Estimate \(c\) and \(\lambda\) with linear regression.

Distance to the mean

Weak stationarity means that the process \(X\) has a fixed mean: $$ E[X] = \mu $$ Taking the expected value of \(X_t\) we get: $$ \begin{aligned} E[X_t] &= E[c + \lambda X_{t-1} + \epsilon_t] \\ \mu &= c + \lambda E[X_{t-1}] \\ \mu &= c + \lambda\mu \\ \mu &= \frac{c}{1 - \lambda} \\ \end{aligned} $$ Rearranging for \(c\) gives: $$ c = \mu(1 - \lambda) $$

Putting this into \((\ref{1})\) gives: $$ \begin{aligned} X_t &= \mu(1 - \lambda) + \lambda X_{t-1} + \epsilon_t \\ &= \mu + \lambda(X_{t-1} - \mu) + \epsilon_t \\ \end{aligned} $$

If we set \(Y_t\) to be the distance from the mean \(X_t - \mu\) then: $$ \begin{align} Y_t = \lambda Y_{t-1} + \epsilon_t \label{2} \end{align} $$

Half-life

The half-life is defined as the time \(h\) it takes for \(X_t\) to decay half way to the mean. In other words, the time \(h\) it takes for \(Y_t\) to decay halfway to \(0\). This is written as: $$ E[Y_{t+h}] = \frac{1}{2}Y_t $$ Equation \((\ref{2})\) gives us: $$ E[Y_{t+h}] = \lambda^h Y_t $$ Which means that: $$ \lambda^h Y_t = \frac{1}{2} Y_t $$

Solving for the half-life \(h\) results in: $$ h = -\frac{\log(2)}{\log(\lambda)} $$

Like what you see? Follow Adrian on Twitter to be notified of new content.

Footnotes

Corrections

If you see any mistakes or room for improvement, please reach out to me on Twitter @DrAdrian.

Citation

Please cite this work as:

Letchford (2023), "Estimating the half-life of a time series", OS Quant.

In BibTeX please use:

@article{Letchford2023,
    author = {Letchford, Adrian},
    title = {Estimating the half-life of a time series},
    journal = {OS Quant},
    year = {2023},
    note = {https://osquant.com/papers/estimating_the_halflife_of_a_time_series/},
}