Hi Moobi,

Thanks for reading! Time series are quite useful because you don’t need outside data to make forecasts. However, they are also limited because they don’t incorporate many of the factors that impact the thing you are forecasting. I’ll try to answer your questions below:

1. There are different ways to measure the accuracy. Prophet has a built in cross validation method. It’s a bit awkward to use but it works. The other way would be to use one of scikitlearns built in error checking methods. The forecast dataframe has the actuals and predicted value, allowing you to calculate R-squared, RMSE, MAE, etc. This post has some details and sample code.
2. Neural networks have been shown lately to outperform time series if there is enough data points. However I don’t believe we are close enough to that with this dataset. Prophet is a simple way of doing time series, you could look in to ARIMA or other time series methods if you want a more statistically d developed technique. I wouldn’t recommend reinforcement learning, but hey who knows?!
3. For learning deep learning I recommend going through a course to learn tensorflow or even better, check out fast.ai. They have a course that can take you from a practical coder to someone who can make usable deep learning models in 7 weeks.
4. Scaling and transformations allow the machine learning algorithm to not over/underweight your features when training. For example, if you have one feature in the 1,000’s range and one in small decimals (.05, etc.) some algorithms don’t handle this well. So normalization or transformations help make your data more standard for the algorithm. Sometimes it helps, sometimes it doesn’t. Dig into the algorithm documentation to learn more.
5. Supervised learning is best for stock prediction. But that’s me personally. Reinforcement learning may be better in a few years but now it seems overly difficult to get usable results.