Bitcoin is the first asset class derived from nascent blockchain technology. Its phenomenal growth in recent years has attracted many investors who used to invest in traditional financial products to start investing in crypto assets. Reports have shown that there is a significant amount of capital flowing into the crypto market. This capital flow has created an increasing demand for new financial products that cater to the needs of the crypto market.
Qidax invested heavily in its Research and Development team to maximise profit through Quant trading. One of the key research area is the development of AI tool in predicting cryptocurrency price like Bitcoin. This article features Artificial Neural Network (ANN) with an “optimised operator” used to correctly predict Bitcoin price.
What is ANN?
The Artificial Neural Network (ANN) is a methodology based on the suspected workings of neurons envisioned back in the 1960s. The era of deep learning has meant significant development of neural nets that are able to learn and distinguish features which are inherently harder for humans to identify. The neural network functionality can best be described as a multiplication of logistic regression modules which enables the network to perform and make inferences on the data presented with much more accuracy and depth than the optimisation algorithms such as random forest and logistic regression.
Artificial neural nets consist of neurons fully connected to each other. As in each neuron is connected to the next layer’s neurons as shown in the figure below.
There are 3 distinct layer types in an ANN. The input, hidden and output layers. The input layers are defined by the number of features in each of the training examples and the output layer comprises of the number of outputs or classifications required. For Bitcoin, Qidax applies symmetric volatility structure of cryptocurrency which can be measured through four input attributes such as open price, high price, low price and close price for predicting its price future trend. A closer look at the Bitcoin prices (consisted of Open, High, Low and Close) recently can be seen in the historical Bitcoin prices figure below:
Next, a sigmoid function is applied to the vector representing the internal weights of the neurons. A more efficient ‘Relu’ or rectified linear unit is used to propagate through the network in order to calculate the outputs layer.
A more efficient ‘Relu’ or rectified linear unit which requires less computational load is adopted in Qidax’s AI model. The rectified linear unit is 0 for all negative values and x for all positive values, represented as python code ‘Relu(x) = max(0,x)’. The sigmoid function is still employed, however it is employed on the final layer to output a 1 or 0.
To optimise the neural net, the gradient descent algorithm is applied to the layers. It is used to calculate the derivatives of the cost function with respect to the parameters W and b. The gradient descent algorithm attempts to find a minimisation of the cost function through differentiating with respect to all parameters W & b in order to back propagate.
Through an extensive period of R&D, Qidax has successfully developed and implemented an efficient AI model in Bitcoin price prediction. Qidax quarterly PnL financial update has proven that ANN is an effective and adequate model with accuracy level of over 80% against actual price. Qidax’s ANN AI tool is deemed a valuable and lucrative investment tool. Not too far down the road, Qidax has a schedule in plan to release a user friendly AI tools to its investors.
Leave a Reply