Cross Validated is a question and answer site for people interested in statistics, machine learning, data analysis, data mining, and data visualization. Join them; it only takes a minute:

Sign up
Here's how it works:
  1. Anybody can ask a question
  2. Anybody can answer
  3. The best answers are voted up and rise to the top

I've worked with certain causal/predictive techniques when handling two time series, but this problem is different from what I am used to and I'm not sure how to proceed.

I would like to see the causal influence of a single event on multiple time series. For example, an earthquake on a particular day having an influence on future stock prices of different companies.

Is there a technique/method that allows me to address this type of question?

share|improve this question

Causal Events ( a.k.a as KNOWN Intervention Series ) can be incorporated into a Transfer Function which might normally include ARIMA structure and also the impact of unspecified (unknown Intervention Series .. but identifiable via Intervention Detection schemes). Software to do this , possible incorporating parameter changes and/or variance changes are not available in the "free software domain" . One can use SAS or SPSS to attempt to construct models requiring you to be an expert in this area of modelling. In terms of full disclosure this is an area that I have specializied in for many years. I suggest that you review my posts on stackexchange for more discussions on this.

share|improve this answer

I think there are several approaches to this:

1) Latent variable growth models: A psychometric approach that uses the concept of "latent variables" (in this case, most probably some idea of "confidence") that explains some extent of variability that is introduced by the discrete event and the subsequent trends in stock, since these trends are easily conceptualized as "self-effecting" (that is perturbations in variance can cause long term ascending or descending trends).

2) Regression analysis: There are a number of approaches here: lagged effects, mixed modeling with ARIMA, cumulative effects, etc. etc. Understanding specification of covariance structure in mixed models is essential to capturing ongoing trends in the outcome.

3) Granger causality: similar to the latent variable approach, this is a method of causal inference similar to a dag approach that tries to isolate specific causal pathways.

4) Marginal Structural Models can be deployed when you are longitudinally measuring other markers for a company's performance over time, such as staffing turnover and layoffs, etc. This is because some exposures vary with time and intensity and their lasting affects modify exposures causing a "cascading" effect in the outcome which may be biased.

Personally, I favor 2. I think that the naive approach for detecting a difference in means in a carefully adjusted regression model, especially in the presence of a well identified control group (such as companies of similar stature in the same country not affected by the earthquake) will provide: 1-easily interprettable effects, 2-high power, and 3-greater generalizability. What some may call a bias in estimates of efficacy in-vitro (such as foolish business leadership mistakes causing a decrease in profitability after the earthquake), are actually unbiased in estimates of effectiveness or in-vivo, because you can't tell whether the foolish business decision would have been made had the earthquake not happened. This highlights the importance of having a well identified control group in any analysis you intend to do.

share|improve this answer

Your Answer


By posting your answer, you agree to the privacy policy and terms of service.

Not the answer you're looking for? Browse other questions tagged or ask your own question.