The trap that matters here. Standard k-fold cross-validation is invalid for time series - random folds train on future data to predict the past.
Use walk-forward validation: train on a window, test on the period immediately following, roll forward. Leave a gap between train and test when features have look-ahead spans, or information leaks across the boundary.
The deeper problem. Repeatedly checking cross-validated performance and adjusting the model makes the estimate optimistic. The only genuinely clean estimate comes from data you have looked at once.