Time series forecasts for Bitcoin: a first draft of our approach

time-series bitcoin

Posted April 30, 2024 by

---

In the previous post, we had to find out that the bitcoin time series is essentially a random walk.

This means, that we can not rely on the bitcoin series itself in oder to make a prediction, at least not on the the bitcoine data alone. And this is just the difference between univariate and multivariate forecasting: while a univariate time series depends only on a single value, a multivariate time series depends on multiple values. At least we have to hope that treating the bitcoin series as a multivariate time series gives us the chance to predict the course.

What could these additional values be? And which methods are we going to use?

Here is a rough plan:

• The data: we will use the exchange rates of the so-called alt-coins. As we could see in the previous post, these often show a similar trend as the bitcoin rate. With a bit of luck, they might be a bit more dynamic than the bitcoin itself and pick up trends earlier.
• The method: we will stick to - at least at the beginning - relatively simple deep learning architectures.
  • We will start with simple and deep RNNs (Recurrent Neural Networks) and LTMS network (Long Term-Short Memory).
  • Later on we might apply more recent architectures (like TSMixer for instance) and more traditional statistical methods.
• The strategy: there are to different approaches that come to mind dealing with data like this:
  • we can tackle the problem by trying to predict the precise course of the btc price in the future.
  • or we can restrict ourselves to predicting the course of the price for the next day. And the prediction doesn't even have to be very precise: it suffices to predict whether the price will go up or down, if we want to benefit financially.

This is really just a rough strategy, but we'll use it as a starting point.