Several people have tried to use technical analysis and chart patterns to predict bitcoin’s future price. These generally use past data and machine learning algorithms or statistical regression analysis in order to try to formulate a general idea of how bitcoin’s price will move in the short term. Although several of these methods exist, I examined the three main ones. This page is a little more technical, so if you are looking for more general trends, please refer to either the homepage or the “Long-Term Survival” page.
The simplest idea is known as the spot price. It assumes that all uncertainty and expectations of future price have already been factored into the price, and thus expects the price to remain at the same level in the near future. For those who are mathematically oriented, essentially the spot price model views BTC price as a martingale (although in the long-term it would be expected to have an upward drift equal to the risk-free rate). For those who are more financially oriented, this view is essentially an equivalent expression to the semi-strong version of the Efficient Market Hypothesis (EMH). Obviously, the biggest problem with this model is that it is trivial, and gives no insight into whether to buy or sell BTC, since it is always ‘fairly valued’.
The second short-term predictive technique I examined was presented in a paper by Shah and Zhang. The paper used nonparametric regression (a method of statistical regression that both determines both the type of function and the function coefficients), and claimed to have returned 89% under its test period. What it did not state, however, was that over the same time (of only 60 days) had someone just bought and held a bitcoin, they would have gained over 100%. Proponents of the paper still argue that it instead provides a hedge, reducing the volatility associated with Bitcoin; however, because it has not been tested in down markets, this cannot be shown. In addition, the paper did not take into account transaction fees associated with buying and selling BTC, which could further have reduced its stated gain.
Nonparametric Regression leads to different test functions along with curve-fitting.
The final method I investigated was Artificial Neural Networks (ANN). ANNs are used for pattern recognition by simulating the way the brain works. Different “neurons” are created that take an input and put in an output. Between the input and output layers of neurons are different “hidden layers” where more non-parametric computation can take place. One of the major problems with Artificial Neural Networks (ANNs) is the question of how many hidden layers to provide, since too many hidden layers can lead to an “overfitting” curve that thinks noise is a part of the general pattern, and too few hidden layers can lead to an “underfiting” curve that ignores some real data. One solution that I hope to implement in the coming year is to create a Genetic Algorithm to optimize the number of hidden layers in the Artificial Neural Network.