I implemented a neural network trained on historical bike sharing data to predict how many bikes would be needed in the future. Check out the full notebook
Hyperparameters:
- iterations = 4000
- learning_rate = 0.75
- hidden_nodes = 18
- output_nodes = 1
Performance: Training loss: 0.065, Validation loss: 0.164
Graphing the Model:
Final Thoughts:
Predictions for December 11 - Dec 21 were close to the actual data. After Dec 21 the model overestimated bike ridership, most likely because it hadn't seen holiday season training examples.