二元变量互相关（时滞相关）：传统互相关函数适用于稀疏时序吗？

Great question—this comes up all the time when working with sparse event data paired with continuous time series, so let’s walk through whether traditional cross-correlation functions (CCF) hold up here, and what to use if they don’t.

First, remember that the standard cross-correlation function computes the Pearson correlation coefficient between two series at every possible lag. Pearson correlation relies on assumptions like joint normality and stable variance—assumptions that get completely thrown out the window with your extremely sparse binary sequence (mostly 0s, rare 1s).

Here’s why it falls short:

• Distribution mismatch: Pearson correlation works best when both variables are roughly normally distributed. Your binary sequence is highly skewed, and its "distribution" is little more than a rare event indicator. This biases the correlation coefficient, making it unable to reflect true lagged relationships.
• Signal drowning: With almost all 0s in the binary series, most lag windows will be dominated by these non-event points. The rare 1s (the actual signals you care about) get lost in the noise of 0s, so the CCF will mostly show near-zero correlation even if there’s a real relationship between the 1s and the continuous series.
• Unstable variance: The variance of a binary variable is (p(1-p)), where (p) is the proportion of 1s. When (p) is tiny (like in your case), this variance approaches 0. This makes the Pearson correlation calculation unstable—small fluctuations in the data can lead to huge, meaningless swings in the coefficient.

Instead of forcing traditional CCF to work, use methods tailored to event-driven sparse data:

• Event-focused correlation: Isolate the time points where the binary sequence is 1, then extract the corresponding (and lagged) values from the continuous series. Compare these values to the continuous series’ baseline (when the binary is 0) using t-tests, Mann-Whitney U tests, or even just mean/median differences. This cuts through the noise of the 0s to focus on the events that matter.
• Lagged conditional probability: Calculate the probability that the binary sequence hits 1 k steps after the continuous series crosses a threshold (or hits a certain range). Or reverse it: compute how the continuous series changes k steps after a 1 appears in the binary sequence. This aligns with the "event" nature of your binary data far better than correlation.
• Robust correlation variants: Swap Pearson correlation for Spearman’s rank correlation in the CCF calculation. Rank correlation doesn’t care about distribution or skewness—it only looks at ordinal relationships, making it much more resilient to sparse binary data. You could also weight the binary sequence, assigning higher importance to the rare 1s to avoid them being drowned out.
• Point process models: Treat your binary sequence as a point process (e.g., a Poisson process) and use coupled models to link it to the continuous series. Models like mutually exciting point processes are designed explicitly to capture lagged relationships between sparse event data and continuous signals, and they come with rigorous statistical guarantees.

Traditional CCF isn’t completely useless here, but its results will be unreliable at best and misleading at worst. You’ll get far more meaningful insights by using methods that account for the sparse, event-driven nature of your binary sequence instead of forcing it into a framework built for continuous, normally distributed data.

内容的提问来源于stack exchange，提问作者Xiaoyu Lu