AI, Machine Learning, and Deep Learning: Business Implications

The hype surrounding Data Science buzzwords like Artificial Intelligence, Machine Learning and Deep Learning is deafening. Every business understands that they have to leverage this to stay competitive. But the way these technologies are often explained, it's hard to understand unless you have a Ph.D.Our experience at Ideas2IT delivering complex Data Science projects for both enterprises and startups taught us that it is important for business stakeholders to have a basic understanding of Machine Learning and Deep Learning. This helps them identify appropriate use cases and to have conversations with the Data Science team.Data Science is a delicate balancing act between Tech, Math/Statistics, and Business Strategy.In building wide-ranging models such as delinquency prediction algorithms and credit scoring engines, we’ve communicated extensively with stakeholders across these functions. Leveraging this experience, we’re going to try and break down some oft-repeated terms related to Data Science, and what they can do for the Finance industry.

First things first: Why should financial institutions care about Data Science?

Today, records in financial services firms are captured electronically by computers connected to the internet. This means that analysts and investors have immediate access to a diverse range of relevant market data in real-time. For example, the online prices of millions of products can be used to assess inflation, movement of scrips in real-time can be used to evaluate and predict general market sentiments, public company data can be leveraged to analyze micro and macro trends. Given the massive amounts of data available today, analysts can (at least in theory) get access to near real-time macro or company-specific data previously not available from traditional methods. In practice, though, a lot of data so gathered isn’t usable as-is. That’s where Data Science and Machine Learning come in handy - To analyze large or unstructured datasets in order to extract actionable signals. This is invaluable and needs to be leveraged today. Already, every financial institution in America is scrambling to attract top Data Science talent. Here's a diagram from JP Morgan illustrating the potential of AI in trading:

Getting down to the definitions:

We’re going to break down some of the popular Data Science related terms in increasing order of sophistication, and also talk about how they could be applied in the Financial industry - we’re taking the use case of rating agencies, but these can be applied across other functions as well - fraud detection, for example, or quantitative trading.

Statistical Modeling:

What it means:

At its core, statistical modeling is a way to use mathematics to find relationships between variables and predict outcomes. This involves 1. A way to create a simplified, mathematically defined approximation of a real-life situation, and 2. A way to make predictions from this approximation. The process of creating this mathematical reality is called modeling, and the set of equations used to arrive at the approximation is the statistical model.

Practical application:

To illustrate how statistical modeling works in Finance, let’s take a specific example: Using Discriminant Analysis for credit rating. Discriminant Analysis involves creating a model to determine which variables are different between two or more groups (description) and classifying any new entities into available groups based on the model (prediction).With Discriminant Analysis, a model can be created with variables such as Country Risk, Industry Risk and Competitive Position, to separate groups into increasing order of risk. This is the description phase. Then, when a new company is introduced, the same model can be used to predict its rating.

Machine Learning:

What it means:

Machine Learning is the application of algorithms to data, in order to predict outcomes. A key difference is that Statistical modeling depends heavily on how the data was collected, the methodology, and the statistical properties of the estimator. Statistics is about drawing valid conclusions, which means that there is a rigorous methodology involved. This tends to restrict the scope of statistical modeling as applied to Finance.Machine Learning, on the other hand, is heavily focused on predictions. Instead of an analyst or programmer understanding the problem well enough to write a program for it, huge amounts of sample data are collected, and the model performs tasks by ‘learning’ from said data. Like William Chen (Data Scientist at Quora) says in this post, “Prediction, performance, and decision-making is king, and the algorithm is only a means to an end."

Practical Application:

Machine Learning when applied to credit ratings can help improve the accuracy of rating predictions, beyond traditional models that employ classical econometrics methods. Using techniques such as K-Nearest Neighbors, Linear Regression and Gradient Boosted Trees, and with relevant heteroscedasticity adjustments, Machine Learning can be used to reduce mean square errors in predictions to within one notch. That is, if the predicted outcome is an AA rating, then the true rating will be between AA- and AA+. This allows financial firms to not merely flag low-rated companies, but come up with probabilistic estimates across segments so that the firms can look at rating in a less rule-based and more continuous manner. Besides, with enough training, cost-effective ML models can even be used to automate corporate credit rating for companies that pass certain basic rating requirements.

Neural Networks:

What it means:

Neural Nets are built similarly to the way neurons are arranged in the human brain. A brain has neurons that are either active or inactive, and synapses that link neurons. With neural nets, neurons are replaced by nodes, and the synapses by edges that represent specific data (like in graphs). Edges are randomly assigned weights, and nodes are given input values from available data. There are “layers” of these node-edge combinations, one on top of the other. These layers together constitute a kind of hierarchy: each layer might represent different parts of a complex problem, or have different ways of evaluating an answer.First, data is fed into the input layer. The “neurons” in that layer are now active, and they have an effect on the weights of the neurons in the next layer, which in turn affect the next layer, and so forth. Eventually, information is passed to the output layer, and the network makes a prediction. If that prediction is right, it is used, and if it isn’t, then data is basically run through the network in reverse (called error backpropagation), and the weights are continuously tweaked until the network is ‘trained’ to make correct predictions.

Practical application:

Neural Network based algorithms generally fare well when lots and lots of data is available, but these models are so complex, they generally become a black box and lose the explainability that simpler models provide (for example, it is hard to understand why the model assigned a particular rating to a company). That’s where Deep Learning comes in. In the case of machine Learning, the algorithm needs to be told how to make an accurate prediction by providing it with more information, whereas, in the case of Deep Learning, the algorithm is able to learn that through its own data processing. It is similar to how a human being would identify something, think about it, and then draw any kind of conclusion.

Deep Learning:

What it means:

Deep Learning is essentially a multi-layer neural network that is employed to solve very complex computational problems. With advances in hardware, training techniques and available data, it has become possible to train neural nets with over five hidden layers - something that was impossible before.

Practical application:

In the case of our example - credit rating - instead of determining the scenarios of low ratings in conventional ML, the Deep Learning approach will be able to sift through the various instances of low ratings and determine the patterns itself. This can be given a name post-facto by us in terms of the “scenario”, and then used to pre-screen companies as they come.All of this cutting edge technology has exciting applications in Finance. We’re tracking them keenly, and as we work on more projects for other clients, we’ll keep this post updated, as well as write up new ones.

Ideas2IT Team