Model training includes estimating and optimizing lag when you tag a variable for lag optimization during the variable selection step. Model training results then includes lag estimates such as this:

The weights indicate which period is taken into account when applying the independent variable. In this example, 100% of the independent variable is taken into account during the initial period (denoted as '0'), along with 45% of the prior period (denoted as '1'), 20% of the period before that (denoted as '2') and so on. This variable shows a lagging effect taking place for up to 4 periods prior to the current period. These weights are technically known as a convolution window.

If the variable is found to have no lagging effect, its weights will look as follows:

In this case the independent variable only affects the dependent variable during the initial period with no lagging weight thereafter.