Direct answer
To assess correlation changes, you need data that lets you (1) measure the relationship between two time series over time, (2) verify where the data came from, (3) ensure the measurements are comparable across time, and (4) check that the result is not an artifact of bad cleaning, missing data, or unstable assumptions.
At minimum, gather the two series you want to relate, the time stamps and sampling rules, the transformation you will apply (for example, using returns rather than raw levels), and documentation of data quality. Also plan the provenance checks and “failure mode” checks (how results can break when data or conditions shift).
Mechanism and definition: what “correlation changes” means
Correlation change means the estimated statistical relationship between two variables is not constant over time. Practically, you estimate correlation over multiple windows (for example, rolling or by regimes) and observe how the estimate moves.
Two key clarifications help you avoid confusion:
- Use appropriate data scale. Correlation is usually computed on transformed series (commonly changes over time). Using raw price/level series can produce misleading dependence.
- Correlation depends on the method and window. Different sampling frequencies, window lengths, and calculation conventions can change the correlation estimate even if the underlying relationship did not “fundamentally” change.
So your “data needed” list is really a list of ingredients for a consistent experiment: the paired time series, how you turn them into comparable numbers, and how you choose the time windows.
Evidence: what inputs to collect (and how to keep them comparable)
Gather the following inputs, and record the assumptions you make:
- The two time series to compare
- Series A and Series B that you want to test as related.
- For each, include the measurement definition (what it represents), units, and whether it is mid-price, last price, quote, or another derived value.
- Time axis and sampling rules
- Start/end timestamps.
- Sampling frequency (e.g., every minute, hourly, daily).
- Whether timestamps are aligned to the same market clock and handling of gaps.
- The transformation for the calculation
- The formula you apply before computing correlation (for example, using returns over each interval rather than raw levels).
- How you treat missing values: exclude intervals, forward-fill, or drop specific points. The choice can materially affect estimates.
- The correlation estimation setup
- Windowing approach: rolling window or fixed subperiods.
- Window length and overlap (overlap changes the smoothness and statistical dependence between estimates).
- Correlation type (commonly Pearson for linear dependence; a different choice measures dependence differently).
- Data provenance and documentation
- Source system/provider identity.
- Any known corporate actions or contract roll rules (when relevant).
- File/version identifiers and retrieval method.
- Quality metrics and cleaning log
- Percentage of missing timestamps per series.
- Outlier detection method and whether outliers are kept or removed.
- Simple sanity checks: are there sudden jumps caused by data resets rather than market movement?
One worked example (assumptions must be explicit)
Assume you select two series observed at the same sampling frequency, and you compute correlation on interval changes. You choose a fixed window length, then slide it forward in time. A correlation “change” is simply the difference between correlation estimates in two adjacent windows.
To make that interpretable, you must state:
- What counts as an interval change.
- How you handle missing timestamps.
- That both series used the same sampling schedule and transformation. If those are not controlled, observed correlation movement can reflect data processing differences rather than any real change.
Limitations and risks: common failure modes
Correlation changes are especially vulnerable to artifacts and instability. Material limitations include:
- Non-stationarity: correlation is not guaranteed to be stable; even with perfect data, relationships can vary as conditions change.
- Sampling and window effects: different frequencies and window sizes can produce different correlation trajectories.
- Data quality problems: missing data, unsynchronized timestamps, or accidental mixing of different instruments/definitions can shift correlation.
- Outliers and heavy tails: a small number of extreme points can dominate the estimate.
- Misinterpretation of causality: correlation change does not prove a cause.
- Costs and execution realities (if you later translate it into decisions): correlation-based expectations can be distorted when real-world frictions apply.