I am currently learning the use of DWT decomposition in time series forecasting with ML models. However, I have a concern regarding potential data leakage during the wavelet decomposition process.

During the convolution step, wavelet filters may use future data points to compute the wavelet coefficients at different levels. This means that even if we split the data into training and testing sets first and then perform wavelet decomposition separately on each, the test set may still "see into the future" — unless we use a causal wavelet filter (i.e., db1).

Now, I have two main questions: