Bálkar Miðlun has developed an AI system, Neural Network, which automatically trades Bitcoin BTC.
Below is the structure of our system and description of our variables.
The date is standard historical data for Bitcoin BTC / USD market
- The time series includes the market data of Open, High, Low, Close prices and the traded Volume
- The data reach from 2013 to now and is continuously updated.
Number of variables
- 10 abstract methods: Technical indicators, Risk Management indicators, Self learned abstractions
- 20 variables are connected to the abstract methods
The Neural Network contains 25 hidden layers
- Weights and bias variables are ca 2500
Convolutional Neural Networks, Fuzzy logic
- The Neural Network breaks down the data to abstract units.
- The abstract units gives us signal for high or low risk, speculative opportunities etc.
- All buy or sell decisions the Neural Network makes are in real time, on data it has not seen before.
- The Neural Network aims to optimize cumulative returns.
- The variables q, are inputs into the Abstract Methods and are positive integers 0-500
- The variables w, are weights between the Abstract Methods, positive numbers
- The variables b, are part of the Hidden layers and work as a bias and are ± n, i.e. b can activate the neuron even if the sum of weights is zero.
- Where r is rear return for each hour t
- Xt is a vector indicating buy or sell signal, 1 or 0 for all t
- VK is the trading fee, in this case 0.25%
- dX/dt is the change of buy or sell vector X for each hour, i.e. how often we trade
Genetic optimization is used to maximize our optimization function Z
- Thousands of random vectors X are created in the process.
- The most promising vectors, X, are cross breaded or mutated generation after generation to breed a strong candidate that maximizes our returns
- Survival of the fittest is what we aim for, candidates that do not perform are killed off.