This is my second model option for forecasting search interest for ten forms of outdoor recreation in Michigan. This option employs the same nonstationary model as model 1, but replaces the VAR model with a neural network for the stationary model. The hope was that this would allow more flexibility in the impact of weather variables. Also, variables were added for holidays.

Results for the neural network are disappointingly close to the VAR model forecasts. In most cases, the nonstationary model dominates the forecast. Despite providing additional variables to the neural network for holidays and time of year, the neural network doesn’t capture holidays or the impact of snowfall much better than the VAR model.

This result is less than hoped but not astoundingly unexpected. There are only a few hundred observations in the dataset and, despite the theoretical ability of a neural network to fit complex nonlinear features, truth is they aren’t magic.

I have one additional modeling technique planned. In the end, the first two methods have done ok. Once I have one more option, I will likely improve the best option. For instance, for the VAR model I could manually add features for holidays or interactions between weather and specific seasons. In the case of the neural
network, I could reduce the variables in the model so it isn’t trying to fit too many irrelevant features.