BFNET (This is not used currently in the model)

(Backward -Forward Network) is the simplest implementation for learning the short-term impacts.

Approach:

Get the last weeks baseline B -1 from the last week’s actual sales by subtracting last week’s additional impacts (marketing impacts) and dividing out last week’s multiplicative impacts (due to external factors such as weather, price, holiday, unemployment rate etc.) After we bring forward the baseline B-1, we assume that to be B0. (predicted Baseline in current period)

Multiply this week’s impacts due to external factors and add this week’s marketing impacts with the baseline B0 to get this week’s sales predictions y-hat. We get actual y from the sales this week. Calculating MSE loss for adjusting the impacts variables for training model.

For calculating the B1, we need to deduct and divide the effects from y actual of current week. The new baseline is carried as B1 baseline to next period, where we multiply and add the impacts to get yhat in period t1.

We aren’t learning decay parameters (how much marketing impacts future sales) in BFNet.