PCA
Non-stationary data
When dealing with non-stationary data, you can take multiple approaches to PCA
- Difference the data, run PCA, then accumulate the estimated factors (@baiPanicAttackUnit)
- Difference the data, run PCA, apply the loading matrix to linearly detrended data (@barigozziLargedimensionalDynamicFactor2021)
- Linearly detrend the data, run PCA on the covariance matrix of the detrended data, apply the loading matrix to the detrended data (@barigozziInferenceHeavyTailedNonstationary2022)
In any case, you must standardize the data first before running PCA.
References
@covarrubiasGoodBadConcentration2019