Predicting Bitcoin using LPM

trialanderror · Post by **trialanderror** » Sat Feb 23, 2019 4:45 pm

Thought I'd try to get some input from the users here. Currently working on a project to predict whether Bitcoin will move up or down tomorrow (and ideally on shorter timeframes too). I believe it could be good to transform the dependent to a binary variable and then run a probit on it, however, it looks like the distribution of returns is more closely described by a leptokurtic distribution. What should I do?

Also, there does not seem to be any autocorrelation in the series. I am currently looking into data from google trends. I would be very happy if some of you could provide examples of you getting creative with data sources and models when trying to predict asset prices.

Any input on the general feasibility of the project is also appreciated.

nikol · Post by **nikol** » Sat Feb 23, 2019 10:09 pm

Basically, there are two components there: diffusion and jump.

Diffusion is more-or-less predictable.

Jump (strong short time movement caused by large volume) is not.

Imagine some hodler who bought 10000 BTC for 1 pizza back in 2010 ${L_SMILIES_SMILE}$ . Now, this pizza maker dumps it, instantly without telling anyone, such that the market cannot absorb it. Boom.

You will never be able to catch it through all this "information diffusion theory"...

trialanderror · Post by **trialanderror** » Sun Feb 24, 2019 11:43 am

I understand that. I wasn't clear in my initial formulation, I am trying to predict the diffusion. Any ideas or general area I should be looking in to do that in your opinion?

riskPremium · Post by **riskPremium** » Sun Feb 24, 2019 3:46 pm

You might want to do some preprocessing on the data first, whether it's price/volume data or alternative data. Think about your stats course in college.

Also start with simple models especially since you are working on daily data (which is not a very large sample).

trialanderror · Post by **trialanderror** » Sun Feb 24, 2019 8:17 pm

I actually am in college right now haha. But yes I am doing preprocessing to the extent I find it useful. It's just that I have a hard time finding a good model when the data pretty much lacks autocorrelation (looking at ACF and PACF)... Currently trying to fit a probit model as said before. I believe that that is fairly simple?