Fisher Variance Stabilization
Key idea
Fisher’s variance stabilization transformation (usually called the Fisher z-transformation) converts a correlation coefficient into a variable whose variance is approximately constant and whose distribution is closer to normal.
It is mainly used for:
- Statistical inference on correlations
- Confidence intervals for correlation
- Hypothesis testing
- Averaging correlations
- Making correlation estimates comparable
The transformation is especially useful because raw correlations have non-constant variance.
Motivation
Suppose sample correlation is:
\[r\in[-1,1]\]The distribution of $r$ is:
- bounded between $-1$ and $1$
- asymmetric near boundaries
- variance depends on true correlation
This makes inference difficult.
For example:
- $r=0.9$ cannot move upward much
- $r=0.1$ has much larger uncertainty
Fisher’s transformation removes much of this issue.
Definition
Transform correlation $r$ into:
\[z = \frac12 \ln\left( \frac{1+r}{1-r} \right)\]Equivalent form:
\[z=\operatorname{arctanh}(r)\]where:
- $r$ = sample correlation
- $z$ = transformed variable
Why it works
If:
\[r\sim\text{Sample Correlation}\]then approximately:
\[z \sim N\left( \operatorname{arctanh}(\rho), \frac1{n-3} \right)\]where:
- $\rho$ = true population correlation
- $n$ = sample size
Important result:
\[\operatorname{Var}(z) \approx \frac1{n-3}\]Notice:
the variance no longer depends on $\rho$.
This is why it is called a variance stabilization transformation.
Example
Suppose:
\[r=0.80\]Transform:
\[z = \frac12 \ln\left( \frac{1+0.8}{1-0.8} \right) = 1.099\]Now inference becomes easier.
To transform back:
\[r = \tanh(z) = \frac{e^{2z}-1}{e^{2z}+1}\]Confidence Interval Example
Observed:
\[r=0.50,\quad n=100\]Step 1:
Transform:
\[z=0.549\]Step 2:
Compute standard error:
\[SE = \frac1{\sqrt{97}} = 0.102\]95% interval:
\[0.549\pm1.96\times0.102\]Result:
\[[0.349,\ 0.749]\]Step 3:
Apply inverse transform:
\[r\in[0.34,\ 0.63]\]Relationship to Other Variance Stabilization Methods
| Data Type | Transformation |
|---|---|
| Correlation | Fisher z |
| Proportion | Arcsin square root |
| Count (Poisson) | Square root |
| Positive skew | Box–Cox |
| General variables | Yeo–Johnson |
Usage in Quantitative Finance
Fisher transformation appears frequently in quant research:
Information coefficient (IC)
Convert daily IC:
\[IC_t \rightarrow z_t = \operatorname{arctanh}(IC_t)\]before averaging.
Correlation estimation
Stabilize rolling factor correlations.
Covariance shrinkage
Work in transformed correlation space.
Signal research
Compare predictive correlations across periods with different sample sizes.
Typical workflow:
\[\text{Correlation} \rightarrow \text{Fisher z} \rightarrow \text{Average / Test} \rightarrow \text{Inverse transform}\]This is one reason IC statistics are often treated using Fisher transforms rather than averaging raw correlations directly.