This is an update for my project where I forecast Google search interest for 10 forms of outdoor recreation in Michigan. The page for this project is pasted below.

I also made some updates to the hyperparameters and settings of the Long Short Term Memory (LSTM) models. This was to modify the keras syntax to change how states and batches are used during model training and prediction. After more investigation and thinking about the problem, I discovered the default keras settings don’t match what you would necessarily expect from LSTM models.

After these improvements to the LSTM models, I have made changes to my model selection. Now I will use a LSTM for all ten forecasts. Previously, for atving, boating, and camping I used a more traditional statistical modeling approach. This will simplify implementation since I’m now using the same model type for all forecasts. It will also make updates easier. As I discussed in my Northern Michigan search interest project, the machine learning approach allows for more automated model re-estimation and updates. With traditional statistical models, there are many more tests and hurdles that need to be passed. Given that I do this project on my free time and considering that I do not have a team of modeling personnel to work with, simpler implementation and updates is often better.

From here on out, the exact settings for these models is proprietary. However, for anyone trying to do something similar, I’ve listed several sources already in Installment 8. In addition to the sources I mentioned in Installment 8, I also used the following.

dataandoutdoors.com/michigan-outdoor-recreation-installment-8/

www.machinelearningmastery.com/understanding-stateful-lstm-recurrent-neural-networks-python-keras/

www.machinelearningmastery.com/use-different-batch-sizes-training-predicting-python-keras/

Thankfully, I’m making progress on this project. I had computer issues during February but made some progress on automating the implementation during January. After implementation, I will start providing recurring forecasts. I’m hoping this will occur in a couple weeks. My initial output will be as simple as possible. However, I will add more information over time.