Interesting article. Financial prediction especially over short time frames is definitely tough.

Three thoughts/questions:

1. It would be interesting to see a longer out of sample period than the 6 months or so that you have. It would be interesting to train your best model on the first 2 years and just do an expanding window after that (always predicting the next outcome variable out of sample with only data that you had up until that point).
2. Once you do go out of sample (the green line), it looks like you are forced to bet on mean reversion (bet that the blue comes back to the green). Given the fit in sample, I assume the trading strategy was not mean reversion? Also, did you find that your prediction led the actual price? If it does not lead the price, then I am not sure if there is a lot of predictive power in the model.
3. Finally, it would be interesting if you applied the same techniques to predicting the next 1 month’s price movement. If it can predict that at a better than a coin flip rate out of sample, then there may be some alpha to be made.