Understanding Stock Market Prediction Using Artificial Neural Networks and Their Adaptation

Understanding Stock Market Prediction Using Artificial Neural Networks and Their Adaptations

taliTali Soroker is a Financial Analyst at I Know First. She graduated from Northeastern University with a Bachelor degree in Mathematics.

Summary

  • Artificial Neural Networks
  • General Applications
  • Different Types of ANNs
    • Feed-Forward ANNs
    • RBF Neural Networks
    • MLP Neural Networks
    • RNNs
  • Neural Networks and Finance

Artificial Neural Networks

Artificial Neural Networks are computing systems modelled after the structure of the neurons in a human brain. This unique assembly of ‘nodes’ allows for a different kind of computing compared to the centralized processing computers that we use in our daily lives.

stock market prediction using artificial neural networks

ANNs don’t rely on a central processor, rather they are composed of many smaller decentralized processing nodes which are organized into computing layers. These nodes simply compute the weighted sum of the outputs from the nodes in the previous layer. In this way, neural networks can take advantage of parallel processing and non-deterministic computing.

Despite their basic structure, neural networks can have vast numbers of hidden layers between the given input and the network’s output, and can be programmed to perform different tasks.

 General Applications

The structure and learning methods of neural networks make them perfect for solving many tasks that our own home computers aren’t designed to solve. Some of the most revered capabilities of neural networks are image and pattern recognition, time series analysis and prediction, and non-linear computation with immense, unlabeled data sets. Already, researchers are using neural networks to create innovative solutions in a diverse range of fields and industries such as finance and medicine.

Automatic teller machines (ATMs) are using these powerful algorithmic systems to interpret handwriting on personal checks and verify the sums when they are deposited. Advancements in self-driving cars are being made by several companies and some self-driving cars are even out on the streets now for training and testing purposes. In the medical field, image recognition capabilities have improved diagnostics for melanoma and other types of cancers.

stock market prediction using artificial neural networks

Different Types of Neural Networks

The architecture of a neural network can change the function of the system, and there are many variations on the basic parallel structure.

Feed-forward ANNs

Feed-forward neural networks are perhaps the most basic of these complex systems. The nodes, or perceptrons, are organized into layers with an input, an output, and some number of hidden layers. Information moves in only one direction (forward) through the network from layer to layer, giving this neural network its name. Each node is connected only to the nodes in the layers immediately before and after the node that it’s in and has no relation to the other nodes in the same layer.

stock market prediction using artificial neural networks

These ANNs use a method of supervised learning called backpropagation in which the system is presented with inputs, and the generated outputs are compared to the desired outputs. In this way, the weighting of the nodes can be adjusted to achieve optimal results. Eventually this process is stopped and the network is run only in the forward direction.

Radial Basis Function (RBF) Neural Networks

stock market prediction using artificial neural networks

Radial Basis Function (RBF) Neural Networks are a subset of feedforward networks with a few defining characteristics. RBFs are based on the theory of function approximation and have only one hidden layer between the input and the output layers. These networks generally utilize Gaussian activation functions and use Euclidean distances between the input and weights, which are viewed as centers. That is to say that the nodes in the hidden layer perform a non-linear function on the inputs and then the output layer maps these non-linear objects into a new space.

RBF functions are used for function approximation and classification problems. These neural networks can be trained for facial tracking and recognition, computer vision, and robotic control. The training time for RBF networks is relatively short, still, they have similar capabilities to the Multi-Layer Perceptron, or MLP, neural networks.

Multiple Layer Perceptron (MLP) Neural Networks

MLP networks are another form of feed-forward networks, though they are more complex and characteristically have more than one layer. Network training is still done with backpropagation for each layer, but the functions coded into the perceptrons vary from those in RBF networks.

These neural networks are called universal approximators, able to approximate functions with even just one single hidden layer provided the hidden layer contains sufficient perceptrons (e.g. nodes). MLPs are especially useful in deep learning and are sometimes referred to interchangeably as Deep Learning Neural Networks. MLPs are particularly useful in evaluating noisy data and classifying untrained patterns.

stock market prediction using artificial neural networks

Recurrent Neural Networks (RNN)

RNNs, or Recurrent Neural Networks, are not feed-forward networks. Rather, these neural networks are more complex in that previous inputs are ‘remembered’ by the system and can be used to gain deeper understanding of sequential data. More simply, when a node in a hidden layer produces an output, this output can cycle back to the same node. This allows the system to learn and identify both spatial and temporal patterns in the original data.

stock market prediction using artificial neural networks

While training for RNNs is considered by most to be more complicated and time-intensive, the applications of Recurrent Neural Networks are invaluable for the analysis of real-world data. Some areas where RNNs have contributed to huge advancements include image recognition and description generation, NLP (Natural Language Processing) and, more lucratively, investment and stock pricing prediction.

Neural Networks and Finance

In the field of economics, much research in recent years has been focused on developing techniques for financial market prediction and forecasting. To predict future stock prices and market fluctuations is not an easy task, considering the innumerous variables that can affect market behavior.

I Know First is one company that is using time-series analysis with deep learning neural networks to do this kind of market analysis and forecasting. The company’s CTO Lipa Roitman has more than 20 years of experience in researching artificial intelligence and machine learning. He is leading I Know First’s R&D team in developing this cutting-edge, deep learning approach to forecasting prices in a chaotic market. The continued efforts of this company helps their investment clients gain insight into potential investment opportunities and generate trading strategies.

The system performs stock market prediction using artificial neural networks that are self-learning, flexible, and adaptive to the capital markets. It identifies market patterns based on data going back more than 10 years which it then uses to produce forecasts for six different time horizons ranging from 3 days to a full year. Each day, as the market changes, the system learns from its past predictions and adjusts the weighting of the hidden nodes to improve its predictive capabilities. In this way, I Know First is able to create wealth management solutions which can substantially outperform the S&P 500’s year-over-year earnings.


premium