Bitcoin neural network analysis

This is not financial advice.

At this point my networks have become quite solid and are producing stable results reliably.
The 1-day network uses data from the past month to predict next day value and is self-compressing the data autonomous. It is tested against a second network where I preprocess the data instead. Both networks generally agree on the trends which is a good sign.
The network improves prediction accuracy versus assuming a flat past growth by 0.032%, which is tested on data not used in training. This may seem tiny, but is comparable to the difference in prediction accuracy between assuming a constant price versus a growing price based on averaged past growth (+0.021% in prediction accuracy [Do not confuse this with 0.021% growth!]).

The 7-day network improves accuracy by 0.045% compared to a fixed growing strategy.

For the one month network there is a philosophical question left. On these scales the data limitation becomes quite severe. Using data from 3 years means there are only 36 fully independent sets of data available. If number of neurons in a layer is comparable, the network can fully absorb all the information. This is not an overtraining issue, it is just the question if you are willing to believe if the future will show a pattern only because it happened like this before just ONCE.
On one week or 1 day there is just so much more data that all patterns are tested against many independent events in the past.

So for now I have drastically restricted the number of neurons of the 1 month network, forcing it to focus only on the most dominating features. This is a very conservative approach yielding an improvement of 0.081%, but this can be improved massively by allowing for more neurons. Not sure yet, maybe I'l end up making two predictions.

These are the predictions. The black line is today 0:00 UTC. Green dots and blue line is the one-day network, blue dots -one week prediction and orange dot one month. Note that the blue line quickly losses accuracy as the errors are multiplied in a repeated one-day analysis. Basically dont trust it after the end of the green dots. Also note that the lines are very flat as the network does not generate fake noise. It looks a bit boring, but everything else would be stupid. I spent a lot of time teaching the network how to filter out all the noise :)

For today the network predicts a boring day with target of 10237$. Tomorrow better. Then next week target is 10445$ and next moth 11506$. So the network sees a nice and steady growth, nothing crazy but above average.