We are forecasting a trend that has been increasing over the last few years with some seasonality effects. The prophet by itself does a good job forecasting the trend and seasonalities. There is a independent variable that has significant impact on our dependent variable. In our historical data there was a period of significant reduction in our dependent variable which was also seen in our independent variable. The relationship is nearly 1:1. However, later in the time series there are reductions that are not driven by our independent variable, but are seasonalities. When adding this independent variable as an add.regressor() it has very marginal effect on the forecast. We have tested 10% to 50% reduction in the independent variable and the response has only changed a few percentage points for the dependent variable. We have tried many of the hyperparameters like the scaling, mode, normalizing the data ourself, etc. Nothing seemed to work except when we log normalized the dependent variable itself (which does not make sense to me why this would have a significant impact). See graph below. Blue line is the dependent and orange line is the independent (note: we are forecasting at the daily level... this graph is at a much higher granularity).
We are forecasting a trend that has been increasing over the last few years with some seasonality effects. The prophet by itself does a good job forecasting the trend and seasonalities. There is a independent variable that has significant impact on our dependent variable. In our historical data there was a period of significant reduction in our dependent variable which was also seen in our independent variable. The relationship is nearly 1:1. However, later in the time series there are reductions that are not driven by our independent variable, but are seasonalities. When adding this independent variable as an add.regressor() it has very marginal effect on the forecast. We have tested 10% to 50% reduction in the independent variable and the response has only changed a few percentage points for the dependent variable. We have tried many of the hyperparameters like the scaling, mode, normalizing the data ourself, etc. Nothing seemed to work except when we log normalized the dependent variable itself (which does not make sense to me why this would have a significant impact). See graph below. Blue line is the dependent and orange line is the independent (note: we are forecasting at the daily level... this graph is at a much higher granularity).