Choosing Cointegration Test Windows Around Research Goals and Breaks
Summary
The discussion considers how to choose the observation window for cointegration tests on financial time series. It offers no formal universal rule: a relationship may appear in one sample length and disappear in another, so comparing results across windows can reveal sensitivity. The appropriate period also depends on whether the aim is statistical description or informing a trading decision, and known events that may shock prices can affect the choice.
A second response recommends using the full available history after cleaning the data and checking for structural breaks. If a break is detected, cointegration tests can be run separately across the resulting regimes. The exchange raises the need for explicit criteria when selecting a preferred window but does not prescribe a test protocol or establish that repeated success across windows validates a pair. Window comparisons and break handling should therefore be treated as diagnostic choices, with conclusions dependent on sample period and purpose.
Key ideas
- Cointegration results can vary with the length of the sample window.
- Window selection should reflect whether the analysis is descriptive or supports trading decisions.
- Known price-shock events may influence which periods are appropriate to examine.
- Using the full history can be reasonable after checking and cleaning the data.
- Structural breaks may justify testing separate periods, but the discussion gives no universal selection rule.
Tags
Full text
# Choosing the time-frame to test for cointegration # Choosing the time-frame to test for cointegration Is there a technique to choose the time-frame for a cointegration test (eg Augmented Dickey-Fueller's)? ## Answer by tagoma (score 4, accepted) https://quant.stackexchange.com/a/3230 I have never read about any formal procedure for this. And, I don't remember this issue is even treated in C.Alexander's book Market Risk Analysis, Practical Financial Econometrics that dedicate a whole part to the cointegration of financial time series. One may well find tests for cointegration succeeding (failing) for a certain time frame and failing (succeeding) for a longer (shorter) time length. This may be of interest to test for integration for different time lengths. And, one may argue the more often it succeed the better it is to take advantage of the paired-behavior of your time series. Another way around is to consider first the time frame you wish, depending on the issue at stake (is that pure stats observation? do you want to ground trading decisions on these tests?). Finally, one may wish to choose time frames avoiding known in advance events likely to cause shocks in prices (company earnings season, ....). This may be of interest to test for integration for different time lengths. And, one may argue the more often it succeed the better it is to take advantage of the paired-behavior of your time series. Another way around is to consider first the time frame you wish, depending on the issue at stake (is that pure stats observation? do you want to ground trading decisions on these tests?). Finally, one may wish to choose time frames avoiding known in advance events likely to cause shocks in prices (company earnings season, ....). Another way around is to consider first the time frame you wish, depending on the issue at stake (is that pure stats observation? do you want to ground trading decisions on these tests?). Finally, one may wish to choose time frames avoiding known in advance events likely to cause shocks in prices (company earnings season, ....). Finally, one may wish to choose time frames avoiding known in advance events likely to cause shocks in prices (company earnings season, ....). ## Answer by Piroinno (score 3) https://quant.stackexchange.com/a/3231 Why not use the entire data. But before you do clean the data by checking for structural breaks (t / F statistic of a dummy variable). If there exist a functional break, then you know you have to perform the co-integration tests separately for each time frames. If you use a preferable time then the question is what criteria will you be choosing this time-frame on.
Shown in full with attribution under the source's licence. Licence: CC BY-SA 4.0 (Stack Exchange)
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.