Short answer: it’s garbage data. I have no confidence at all that it is correct. Or, more accurately, I think it’s probably correct—since 100 out of the 116 results in the source were ‘Full Shadow’—but only for the same reason that a clock is right twice a day.
The size of the dataset is small but that’s not necessarily invalidating. Small datasets can be fine if the scope of the problem is small. I’d say the features are more important. In this case, the features are the average temperatures for February and March which were in the original dataset. These are attempting to measure how much “winter” there was over those six weeks. This is a reasonable way to measure winter.
In this exercise, I was turning the classic weather prediction method on its head. Rather than predict the weather using Phil, I would predict Phil using the weather forecast. Phil is wrong the majority of the time, and the data reflects that. And this inaccurate data was used to train the model. So, the model has the same biases that the data does. In this case, Phil’s reaction is, essential, random. So, the accuracy of the model is terrible.
In other words: Garbage In. Garbage Out.
As far as the Postman stuff goes, Postman is basically just a GUI version of curl. I’m using it to interact with our API because it’s quick, easy, and I can share the calls I make with it. We’ve got some tutorials on getting postman setup that you can check out.
Once you have Postman installed and have an API key, you can import the collection I shared on GitHub and try running the model yourself.