Artificial Neural Networks for Finance

Back in the early days of data science, before it was even called data science, any financial applications handled by programs were called Expert Systems. These were a domain of AI that was developed using the knowledge of a “Human Expert.” The expert’s knowledge was used to create a set of programming rules to assist the algorithm with making decisions. 

At its most basic level, an Expert System would look like this:

If the price of asset “A” when compared to asset “B” exceeds X%, then sell asset A (or buy asset B or do both), or: 

If a prospective borrower has a credit score below 591, do not lend them anything.

Such expert systems have been successfully used in fraud detection, medical diagnosis, and even when prospecting for minerals. However, there is a major limitation to them, which is that they require full information to be provided to them as an input and this fact means that they will either perform poorly or not at all with uncertainty. 

Financial applications primarily deal with the prediction of future events based on the results of past data. This is the reason that Artificial Neural Networks have become so popular in recent times, especially in the finance industry, because they have a better ability to handle uncertainty when compared to expert systems. When we consider various scenarios that involve predictions, we find a few primary areas enhanced by using artificial neural networks (ANNs): 

1.     Predicting the movement of the stock market, both indexes, and individual stocks

2.     Predicting loan application underwriting and repayment success

3.     Finding suitable credit card clients

In this article, we will explain the basics of artificial neural networks and go deeper into the applications where artificial neural networks can be the most successful and beneficial for the financial, banking, and insurance industries. Finally, we will finish with an example outline of an ANN for making credit decisions.

Artificial Neural Networks in Brief

ANNs are designed to mimic the actions of biological neural networks seen in life forms with nervous systems and brains such as humans.  

Image courtesy of Quora

The biologic nerve cell will take the chemical input into its dendrites, and if the signal is sufficient, then it will transfer this signal down its axon to its axon terminals, where it produces its own chemical signal to go to the next nerve cell.  

The artificial neuron (sometimes called a perceptron) will take input and evaluate it with a bias (or summing) function. The bias function decides what to do with the result, sending it on or not, and to what degree the message will be transferred.  

This perceptron was created by Frank Rosenblatt back in the 1950s and was used by the US Navy for image recognition tasks as well as many other applications.  

The ANN expands on the perceptron and consists of many interconnected neurons all performing their summing functions with the data inputs. Each of the following circles is a single artificial neuron.

Image courtesy of techvidvan.com

The ANN is made up of input and output layers, and a network will have at least one hidden layer between these, but can have dozens of hidden layers with numerous neurons in each layer depending on the model. 

The summing functions for each neuron (colored circles above) will have their own weights and use input data coming in from the left and are connected to the next layer to the right, where they send their decision results. Information is stored in the weights of the connections between the neurons. As an ANN is “trained,” the weights are what changes to improve the results that the model is providing as its output. 

This example is a “feed-forward architecture” and the most commonly used in ANN applications. There are other types of neural networks that are used in specific applications where they perform better. 

ANNs give the user the ability to utilize the data available fully and to determine the structure and parameters of a model without restrictive modeling assumptions.  

Artificial Neural Network Applications

ANNs are especially appealing in finance, banking, and insurance because there is an abundance of high-quality data available for these fields. This data means that there are plenty of inputs, and before ANNs, a lack of testable financial models to deal with all this data.  

Predicting Stock Movements

The prediction of stock market indices and specific stock values are handled by ANN using the vast supply of historical data and then predicting based on several parameters. The accuracy of the prediction is enhanced by the choice of the variables and the information that is provided during the training process.

It can be further improved with an ANN structure that has more hidden layers and more training variables. One group attempted to predict the NASDAQ stock exchange movement and found that a network with three hidden layers, consisting of a configuration of 20-40-20 neurons in the hidden layers, the team had an optimized network and a resulting accuracy of 94.08%. 

While there are other types of neural networks, these types of feed-forward networks are the most widely used because they offer generalization abilities and can be implemented easily.   

Searching for Credit Card Customers

Some credit card companies are using ANNs to decide whether to grant credit card applications. The underwriting process uses the analysis of past failures to make current decisions based on the past experience of other card holders.  

All banks that are in the credit card business wish to obtain an ideal customer who will help them remain profitable. If the client does not spend much with their credit card or uses the revolving line of credit, then that customer is not profitable.  This non-profitable customer will have a per card revenue much lower than the per card cost, and the result will be a low breakeven percentage. 

A group of researchers used an artificial neural network to approach this problem and more accurately predict ideal customers. This study used values called eigenvalues to find the lowest error rates for deciding on the best customers. After several rounds of testing, there were 14 eigenvalues that had the lowest error rates identified and settled on when choosing the most suitable customers. 

This process eliminates instances where credit cards are issued to customers who have no credit card needs, and it gives the bank more meaningful questions to ask on a credit application to better identify the ideal customer.  

This is now broadening beyond the yes-no approval decision and expanding to the amount of credit that is being provided to customers who are approved.

Evaluating Loan Applications

Financial institutions will provide loans to their clients for different reasons, and these decisions are based on various factors. ANNs can be employed to aid in the underwriting process, deciding whether to approve or decline the loan application. 

Any loaning institution will want to minimize its default rate for loan applications and maximize its returns on the loans they issue. A research group was able to test the accuracy of an ANN in predicting the success of loan recovery, and they found an accuracy of 92.6%. 

Additionally, their error rates for Type I (making a bad loan) and Type II (rejecting a good loan) errors were 6.5% and 8.2%, respectively. The failure rates that have been seen for loans approved using ANNs are lower than some of the best traditional methods.  

Other Applications

Beyond those applications listed above, ANNs can be applied to several valuable use cases:

·       Forex price predictions

·       Futures movements and pricing

·       Bond ratings

·       Prediction of business failures

·       Assessment of debt risk

·       Predicting bank failure

·       Bank theft

·       Predicting recessions

How an Artificial Neural Network Decisions Works 

To give an example of how an ANN decision can be made, let’s consider an example of what could be used to make a creditworthiness decision.   

Inputs

·       Age

·       Gender

·       Annual Income

·       Length of time at current job

·       Marital Status

·       Number of Children

·       Number of Children in the home

·       Education level attainment

·       Homeowner or renter

·       Cars owned

·       Address/area

·       Commute distance

·       Credit score

Training and Testing

There will be a large set of clean data created that contains all the inputs to be fed into the ANN to train it with known results (this is called the training set), changing the weighted variables for each neural node to increase the model’s prediction accuracy.  

Once the ANN is trained, a different set of input data is supplied (none of which is present in the training set), and the ANNs “Loan Approved” results are obtained. This second run through of data is a “test set” and can be done using real-time data coming in. 

Based on what was “learned” during the model’s training phase, the accuracy of the predicting ability is refined. The model’s prediction accuracy depends on the various input factors that go into it as well as the addition of hidden layers which are added to the neural network–until the optimum level of accuracy is achieved.

Closing Thoughts

ANNs have continued to improve, and their use has broadened with the decreases in computer costs and the persistent increases in computing power. They will likely be a foundation for financial and economic models but may need to evolve with the likely adoption of quantum computers. 

As we move forward, ANNs will become an even more useful tool to automate service- or data-oriented tasks. The financial, banking, and insurance worlds have an abundance of clean data that can feed into ANNs. 

Care must be taken to ensure that bias is removed from any data going into the model as this can ensure that bias will not come out of the model. Essentially, if we treat the models with care, they will bring us infinite value.

Disclaimer: The author of this text, Jean Chalopin, is a global business leader with a background encompassing banking, biotech, and entertainment. Mr. Chalopin is Chairman of Deltec International Group, deltecbankstag.wpengine.com.

The co-author of this text, Robin Trehan, has a bachelor’s degree in economics, a master’s in international business and finance, and an MBA in electronic business. Mr. Trehan is a Senior VP at Deltec International Group, deltecbankstag.wpengine.com.

The views, thoughts, and opinions expressed in this text are solely the views of the authors, and do not necessarily reflect those of Deltec International Group, its subsidiaries, and/or its employees.