Direct answer
Volatility in Pair Correlation means that the relationship between two time series is not stable over time. To measure it, first compute Pair Correlation as a time-varying correlation (for example, using a rolling window). Then measure how much that rolling correlation itself fluctuates—using statistics such as the standard deviation (or other spread measures) of the rolling correlation series.
This approach focuses on measurement of change, not prediction of future direction.
Mechanism or definition
Start with two aligned time series, for example returns from two instruments: xₜ and yₜ. Pair Correlation is usually defined as the correlation between x and y over a selected period.
To make correlation time-varying, use a rolling window of length W:
- For each time t, take the last W observations {xₜ₋W+1…xₜ} and {yₜ₋W+1…yₜ}.
- Compute correlation over that window, producing a sequence rₜ (the “rolling Pair Correlation”).
To measure volatility in that relationship, compute a second-layer measure over {rₜ} across a longer evaluation period:
- Volatility-of-correlation (spread): the standard deviation of rₜ values.
- Alternative spread measures: interquartile range (IQR) of rₜ, or mean absolute deviation.
- Regime-change lens (optional): count how often rₜ crosses user-chosen thresholds (e.g., becomes strongly positive/negative), as a descriptive way to capture instability.
Key assumptions must be explicit:
- You need consistent sampling frequency (same time step for both series).
- You need aligned timestamps (handle missing values consistently).
- You must choose what “returns” or changes mean (e.g., log returns vs simple returns) and apply it consistently.
Evidence or example
Consider a simple example with two return series over N observations. Let W be the rolling window length. You compute rₜ for each t where a full window exists, giving a vector of rolling correlations.
Now suppose you also compute the volatility-of-correlation as the standard deviation of rₜ over the evaluation range. If rₜ stays close to a single value, the standard deviation is small, indicating a relatively stable co-movement. If rₜ swings widely (even if the average correlation is near zero), the standard deviation is larger, indicating higher volatility in the relationship.
You can also compare how results change when you vary W:
- A shorter W can react quickly to local changes but may be noisier.
- A longer W smooths changes but may hide short-lived instability.
Limitations and risks
- Window-size sensitivity (failure mode): The measured volatility depends strongly on W. Different W values can lead to different conclusions about “stability,” especially when relationships shift.
- Correlation is not causation: A changing correlation does not explain why it changes.
- Regime breaks: Correlations are historical. Relationships can change when market regimes shift, liquidity changes, or the nature of price movements changes.
- Data and alignment quality: Missing data, asynchronous sampling, and inconsistent preprocessing (e.g., different return definitions) can create artificial correlation volatility.
- Non-stationarity of variance: If each series has time-varying volatility, correlation estimates can become unstable even without a meaningful change in the underlying dependence structure.
These are reasons to treat the measurement as descriptive and scenario-dependent.
Verification or next question
To verify the measurement choices independently:
- Recompute rolling correlation with at least two plausible window lengths and check whether the volatility-of-correlation conclusion is qualitatively similar.
- Confirm that the two series are aligned to the same time grid and use a single consistent return definition.
- Examine the rolling correlation time series rₜ directly (not only the summary volatility) to distinguish gradual drift from abrupt breaks.
A useful next question is: Which definition best matches the dependence you care about—linear correlation (Pearson) or a different dependence measure that can behave differently under tail events?