Correlation vs causation in trading signals
Posted: Wed Oct 28, 2026 6:12 am
A signal correlates with subsequent returns. Does it cause them? In markets, the distinction matters less than in medicine, since trading needs prediction, not explanation. But the difference still shapes how reliable a signal is likely to be.
If a signal predicts returns because it captures a real mechanism, like order flow imbalance or a known institutional flow, it has a good chance of persisting. If the correlation arises from coincidence or a third factor, it can vanish when the factor changes.
Examples: a rise in a currency pair coinciding with a rise in a particular indicator may simply reflect that both respond to the same underlying news. Trading the indicator in isolation then works only while the news-driven pattern lasts.
Spurious correlations abound in time series data, especially with trending series. Two unrelated trending series will often correlate strongly. Using returns, not price levels, and testing on out-of-sample data reduces the risk.
A good habit is asking "why would this work?" before any test, and "who is on the other side and why would they lose?" before any trade.
The same caution applies to correlation between a trader's own habits and outcomes: Monday trades, coffee, a particular chair. With enough variables, some will correlate by chance. Treat each as a hypothesis to test on new data, not as a discovery.
It helps to keep a short list of pre-stated hypotheses and to review them only after a pre-set number of trades has accumulated. Peeking early is another form of snooping.
What mechanism explains the main trading signal, and what would break it?
If a signal predicts returns because it captures a real mechanism, like order flow imbalance or a known institutional flow, it has a good chance of persisting. If the correlation arises from coincidence or a third factor, it can vanish when the factor changes.
Examples: a rise in a currency pair coinciding with a rise in a particular indicator may simply reflect that both respond to the same underlying news. Trading the indicator in isolation then works only while the news-driven pattern lasts.
Spurious correlations abound in time series data, especially with trending series. Two unrelated trending series will often correlate strongly. Using returns, not price levels, and testing on out-of-sample data reduces the risk.
A good habit is asking "why would this work?" before any test, and "who is on the other side and why would they lose?" before any trade.
The same caution applies to correlation between a trader's own habits and outcomes: Monday trades, coffee, a particular chair. With enough variables, some will correlate by chance. Treat each as a hypothesis to test on new data, not as a discovery.
It helps to keep a short list of pre-stated hypotheses and to review them only after a pre-set number of trades has accumulated. Peeking early is another form of snooping.
What mechanism explains the main trading signal, and what would break it?