A time series model is evaluated using random k-fold cross-validation. What leakage risk does this introduce?