So, how does Neural Tip’s AI work?
Neural Tip, like its name suggests, uses a classic form of artificial intelligence; neural networks. Classed mathematically as “Universal Function Approximators“, Neural Networks are nothing more than functional nodes arranged in layers, with one layer called the Input Layer, one layer called the Output Layer and an arbitrary number of layers in-between called “Hidden Layers“.
Each layer, as mentioned before, has “Nodes” or “Neurons” that act as a simple mathematical operation on a number of inputs. These Inputs for a particular layer can come from the layer before, or if the layer in question is the Input Layer, simply be the raw data inputs fed to the algorithm. For hidden layers, each node in the layer can take inputs from 1 or more nodes in the layer before it.
Each such link, from a node in one layer to a node in another, has an associated “Weight” that multiplies the value of the output from the source node and passes it to the destination node. When this weight multiplied value comes in as input to the destination node, the destination node carries out an “Neural Activation Function” on said value. If it is above a certain threshold, the node/neuron outputs a value close to “1″, if it isn’t it outputs a value close to “0″, some network setups can also have neurons output negative values. This output is again forwarded to the neurons this neuron connects to in the next layer and so on and so on.
Effectively this can model very complex functions. A graph like illustration of this network architecture is shown below:
So you have an idea how the Neural Network processes data and “thinks”, but, how does it learn?
How Neural Networks Learn:
The reason Neural Networks can be so powerful is that they can learn complex models and relationships between data. This learning can be accomplished in several ways, but the simplest is a formula to randomize the “Weights” initially and then iteratively and gradually change them by trying out a “prediction” on an example piece of data. The actual “answer” or desired output is compared with the output you got with just random weights, this comparison is then used to move the weights closer to the “right answer”. Neural Tip uses much more sophisticated architectures, error propagation and optimization methods than this however. Our algorithm is a combination of recurrent and multi-layer “Deep” Neural Networks that take into account many different variables. Some of these include:
Sentiment Analysis Scores of Contextual text in News Articles and “Stock Twit” tweets for a Ticker Symbol
Daily and Trended Volume
Historical Price Action in Time Frames (5 Day Historical, 30 Day Historical etc.)
Earnings Per Share
S&P, DJIA & NASDAQ Historicals