Deep Learning Finance: Artificial Neural Networks, Deep Learning and Applications of Deep Learning in Finance

taliTali Soroker is a Financial Analyst at I Know First. She holds a B.S. Mathematics from Northeastern University

Deep Learning Finance


  • Artificial Neural Networks Overview
  • ANNs Vs. Conventional Computing
  • Structure and Backpropagation
  • Deep Learning
  • Applications in Finance and Other Areas

Deep Learning Finance

Imagine you’re handed a piece of paper with a picture of an animal on it. You know that it’s an animal, and you know which one. Now, try to explain to somebody else how you recognized which animal it was. Which characteristics told you that it was a dog and not a cat? How did you know which characteristics were the most important to identify? If that was too easy for you, think about how you might instruct a computer to distinguish the differences between animals and to identify a given animal from an image.

Billions of neurons in our brains are interconnected to form a complex system of communication. These neurons help the brain communicate with the rest of the body and allow us to process information, and ultimately, to perform this kind of simple object recognition. Based loosely on this biological system, scientists and mathematicians developed Artificial Neural Networks (ANNs) to execute similar tasks to people, and in some cases, with more efficiency.

Artificial Neural Networks

Artificial Neural Networks are most commonly designed in layers with an input, an output, and one or more hidden layers in between. When data is presented to the input layer, weighted calculations that are programmed into each individual ‘neuron,’ or node, direct the information through the system layer by layer until it reaches the output layer.

The key distinction between these neural networks and conventional computers is in their processing style and learning methods. In a conventional computing system, a central processor is given a defined set of rules and it can then call on data in a logical and sequential way. Conventional computers are great for accounting, word processing and mathematical cataloguing. These computers are the ones that we use daily to surf the web and to write articles like this one.

On the other hand, ANNs operate like an assembly of many simple processors rather than one complex central processor. These smaller processors are the interconnected nodes embedded in each layer, and they do little more than compute the weighted sum of inputs from other nodes.  Thus ANNs can take full advantage of parallel processing hardware, such as graphics card or cloud service, which results in much faster learning. Additionally, neural networks don’t follow any given set of rules, rather they learn by example and are excellent at recognizing patterns in speech, images, text, and more.

ANNs are non-deterministic processors, meaning that the output is not fully determined by the input. So, in the simple ANN represented below, Input A could lead to Node B with 70% and to Node C with 30%. Then, we can reach Output D from different paths or we can reach another output entirely. The same input can theoretically lead to a different output OR different inputs can lead to the same output. This is important in image or handwriting recognition in which different-looking images or letters should be classified the same way.

Neural Networks Learning Methods

Perhaps the biggest distinction between ANNs and conventional computers is in how they learn. Neural networks employ what is called backpropagation, or the backward propagation of errors, as a learning mechanism. The backpropagation algorithm is an expression of the change in output relative to a change in weighting within the hidden layers. Essentially, it allows the system to compare its output with the desired output and then adjust its connection weights accordingly.

In this way, the network can continuously improve itself without human interaction until it reaches a level of acceptable accuracy. Once it reaches this point, the network is run in forward propagation mode as an analytical tool. If the system is allowed to continue adjusting the weights, eventually it will become too specified to the particular input and will lose predicting value. When humans make this mistake, it is called “learning by rote”, when ANN does it, it’s called overfitting. To prevent overfitting the ANN is tested periodically on a new data that it hasn’t seen before. When the errors begin to increase, the learning is stopped.

Deep Neural Networks

Within the realm of neural networks, there are more advanced systems called Deep Neural Networks (DNNs). Networks capable of ‘deep learning’ have multiple hidden layers. When learning is passed from one hidden layer to the next, it achieves a higher level of abstraction when approaching tasks. Thus, when one layer recognizes a shape of an ear or a leg, the next layer could tell if it’s a cat or a dog.

Applications of Deep Learning

The applications of deep learning technology are endless, and recently, research about artificial intelligence and deep learning, in particular, has increased dramatically. Many of Google’s developments, including automatic image translation in Google Translate and their much-anticipated self-driving cars, take advantage of the power of deep neural networks.

Beyond advancements in technology, deep learning has progressed the medical, engineering and finance fields among others. Some exciting developments include improved detection of melanoma and brain cancer, predicting the structural response of buildings during an earthquake, and forecasting financial market shifts.

Deep Learning in Finance

In the world of finance, researchers have been looking at many different areas where artificial intelligence could be helpful. This includes analyzing trading strategies, predicting corporate bankruptcy, and examining the overall health of larger banking and financial systems. Artificial Neural Networks are also being used to perform stock predictions and market forecasting. The propensity for ANNs to interpret and find patterns in massive sets of unlabeled data makes them ideal for identifying patterns in market data.

I Know First is one of the companies pioneering the use of this new technology to increase market profits for its clients. The company’s CTO Lipa Roitman developed a predictive system based on genetic algorithms and unsupervised machine learning, or deep learning. Based on 15 years of historical data and current market data, the system can identify patterns and predict future shifts in stock share prices over six different time horizons. Each day, as new data is recorded, the system learns from its successes and failures and adjusts itself accordingly. In this way, I Know First’s predictive system is continuously improving its own results.