Using the Box-Cox Transformation to Adjust Financial Time Series

Summary

The article explains the one-parameter Box-Cox transformation as a way to make a data series’ distribution more compatible with statistical methods that assume normality. The parameter controls a power transformation, with the logarithm as a special case. The article describes choosing the parameter by maximizing a log-likelihood objective, using a search procedure or an optimizer such as Powell’s method. It also outlines an alternative based on correlation with normal-distribution quantiles.

For financial series, the method may help address skewed distributions, but it does not make data stationary and cannot guarantee a normal result. The transformation requires strictly positive inputs, so the article shifts the series before applying it. It further cautions that the best transformation for forecasting may be the one that minimizes forecast error, rather than the one that makes the data look most normal. No empirical trading results are presented; the discussion is methodological and focused on statistical analysis of quote data.

Key ideas

• The Box-Cox family uses a single parameter to apply power transformations, with a logarithm at the zero-parameter case.
• The transformation parameter can be estimated by maximizing a likelihood-based objective.
• Input observations must be positive, so nonpositive series require a shift before transformation.
• A Box-Cox transform can move a distribution closer to normality but cannot guarantee normality or stationarity.
• Forecast error may be a more relevant criterion for choosing a transformation than normality alone.

Tags

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