Data set: response and predictors are all non-stationary, time series variables
After performing Box-Cox transformations and testing a variety of power transformations on each variable, the non-stationarity still exists. Linear models have therefore been discarded in favor of autoregressive models.
To construct ARIMA models, the variables were differenced, but the acf and pacf plots still have bars that exceed the significance bounds at all difference levels.
What statistical approach is next in line? Nonlinear transformations? Nonlinear models?