The analysis of our autopilot system for predicting crypto prices (part 1)
On our forum I have received the following comment:
Based on the last gif where you show how your predictions evolve and the real stats proceed over time it seems like they are not even close to be accurate.
Great question! On first sight it may appear to be inaccurate, but there is a reason for that. Allow me to explain how it works.
Here is the gif @ses is referring to:
A certain frame on this gif is the chart below. It was generated for the interval 22:00 → 22:30 (GMT+1). And it shows that the price is most likely to crash/go down from 11.5k to 10.9k:
But after a few hours its outcome has been adjusted. It learned that the price didn’t crash so the outcome started to change:
But if we look on the screenshot below, and compare it with the prediction above from 06:00 to 12:00 , it appears to be a pretty good prediction of how the price evolved. At first it went up a little in the first few hours and then slowly started going down. But right now (as I am writing this post), it indicates the price may crash/go down again:
Here’s an easier way to look at what I mean:
It’s important to remember that the screenshots above all use 30min intervals. So the neural network is only trained on plots at 30min intervals, it’s not very good at predicting e.g. +8 hours into the future because that’s already 16 intervals it has to predict, while we noticed it’s only good for up to 5–10 intervals into the future. So to tackle this problem you may want to use hourly/12h/daily intervals instead:
Below are two screenshots which use hourly intervals.
The first one makes a prediction on Jan 27, 22:00 for the next 20 hours:
The second one I have pulled from the site right now and shows how the price evolved, including its prediction:
Now we can compare the first prediction with the 2nd screenshot’s actual data. Now find the x-tick “Jan 27, 22:00” on both screenshots and look how the prediction differs from actual data.
We do see that at 02:00 → 06:00 the prediction showed it would go down, the reality was less worse than indicated, we did have a dip at 01:00 and at 04:00.
Then from 06:00 until 14:00 the price went up and both areas look very similar, so the prediction is pretty accurate for this zone.
Here is a better illustration of what I mean:
And finally, for the 12hr interval we have the following (see image below). It predicted the price to go up to 11.5k and so it did in reality.
On first sight it appears that hourly/12h intervals are much more accurate than 30min intervals. We don’t have enough data yet to analyze the 24hr intervals, but in due time we’ll dig deeper.
I hope this clears things up a bit :)
- Ilya Nevolin