Last time, two years ago, I did a simple experiment to check if a six layer convolutional network could predict if BitMEX's next 5 minute candle is closing higher or lower than the previous one. The accuracy I achieved was 55%, i.e. 5% better than chance. I.e. a useless. I.e. ANN's are no good for straightforward analysis?
Wrong. I framed a better question this time around: is the current trend going up or down? As start and endpoint of a trend I choose the highest and lowest price points within N minutes. That question is fairly easily answered by the ANN, this is what the accuracy looks like after 50 minutes of training:
I haven't yet analysed how to use the data, as just trivially placing market orders in the direction indicated by the ANN is hardly the ideal solution due to jitter and what not.
I'll do some quick evaluations of different N (time scales) and then get to work with the analysis of what precisely I can do with this excellent data. Wish me luck!
