Correlation vs causation investing matters because two assets moving together doesn't mean one drives the other.

Two series that move closely together on a chart can look convincing. For a risk manager, however, the key question is not only whether two variables move together, but why the link exists and whether it is likely to survive different market conditions.

That distinction matters when historical correlations inform portfolio assumptions, hedges, forecasts, or risk models.

What is the Difference Between Correlation and Causation in Markets?

Correlation shows whether two variables move together, while causation indicates that changes in one variable contribute to changes in another.

In markets, two assets can move together without either one directly driving the other. Interest rates, inflation expectations, liquidity, growth, or another common factor may influence both.

This is why correlation vs. causation matters for risk managers: if a shared factor changes, a strong historical relationship may no longer hold.

Why Can Market Correlations Look Stable Until They Do Not?

Historical correlation depends on the period, data, and market environment being measured. A pattern observed during one regime may change when monetary policy, volatility, liquidity, or investor behavior shifts.

This creates correlation risk: the possibility that an assumed relationship stops behaving as it has in the past.

For example, two assets may appear to provide diversification benefits in normal markets but become more closely correlated during periods of stress.

For much of the roughly two decades preceding 2021, US stock and government-bond return frequently exhibited negative correlation. During the higher-inflation environment of 2021 and 2022, that link changed, and positive correlation became more prominent. This illustrates how diversification can vary across market regimes.

A historical coefficient reflects only the sample behind it, not future conditions.

What Does a Documented Spurious Correlation Look Like?

An example comes from David Leinweber’s paper on data mining, “Stupid Data Miner Tricks: Overfitting the S&P 500.”

Leinweber showed that regressions using unrelated economic series, including butter production in Bangladesh, could fit S&P 500 movements closely in sample. Adding other unrelated series, including US cheese production and sheep populations, further improved the statistical fit.

The example was intentionally constructed to illustrate the risks of data mining and overfitting, not to suggest that these variables had a genuine causal relationship with the S&P 500.

This is a classic spurious correlation. If enough variables are tested, some may appear to explain market outcomes simply by chance.

It also shows why you should assess statistical fit alongside economic reasoning, robustness testing, and out of sample evidence before incorporating a relationship into a model.

How Can Correlation Become a Risk Model Assumption?

The risk starts when an observed relationship becomes a built-in model assumption.

A portfolio model may assume two exposures will keep offsetting each other’s risk. Similarly, a hedge model may rely on a historical link between an exposure and its hedge. A forecasting model may use a correlated variable as though it were an enduring economic driver.

These assumptions drive position sizing, hedging, risk limits, and scenario estimates.

This illustrates why risk models' underlying assumptions are commonly tested and monitored. A model can be mathematically sound yet produce misleading outputs if its underlying relationships are unstable.

US supervisory guidance on model risk management describes practices such as testing models using data outside the development sample, comparing results against alternative assumptions, assessing data quality, and conducting ongoing monitoring as conditions change.

This guidance is cited only as an example of model risk management practice. It does not imply that US banking rules apply elsewhere.

What Should Risk Managers Check Before Trusting a Correlation?

When assessing an observed correlation, risk managers commonly consider:

1. Is there a plausible mechanism?

   Identify why the variables should move together rather than relying only on statistical fit.

2. Is there a shared driver?

   Test whether rates, inflation, liquidity, volatility, or another factor explains both variables.

3. Does it survive different samples?

   Compare different periods, market regimes, and out of sample results.

4. What happens under stress?

   Consider scenarios in which the relationship weakens, reverse, or disappears.

5. What decision depends on it?

   The more material the model outputs, the more scrutiny the underlying assumptions deserve.

CFA Institute research has also noted that conventional factor models can confuse statistical association with causation, and that a strong fit alone does not prove the model captures the real drivers.

What Does a Correlation Breakdown Mean in Practice?

A correlation breakdown does not necessarily mean the original analysis was useless. It may indicate a market regime change, a weakened common driver, or a relationship that was more conditional than expected.

Ongoing monitoring can help assess whether model assumptions remain appropriate over time, whether hedge effectiveness changes, and whether market conditions move beyond the regimes represented in historical data.

If market conditions change materially, the assumptions may need to be reassessed.

Conclusion

Correlation remains useful for identifying relationships, supporting diversification analysis, and organizing market information.

Historical co movement alone is not sufficient evidence that the same pattern will persist or that one variable cause another.

Assessing correlation alongside its underlying mechanism, market context, and performance across different scenarios can provide a more complete view of the relationship and its limitations.