Tech trends: machine learning and artificial intelligence

July 25, 2016 by Kyle Johnson

about the author:

Kyle Johnson

Revenue operations lead

Kyle provides revenue and sales analysis for all Advicent go-to-market teams. These analytics optimize Advicent pipeline forecasting, marketing strategies, and leveraged media channels to improve efficiency of sales operations. Kyle is interested in combining his three passions: tech, data analytics, and marketing, to drive success.

“All models are wrong, but some are useful,” said George E.P. Box, a famous statistician. Everyone has mental models of how the world works because a model provides a simplified explanation for a very complex world. Some common mental models include Occam’s RazorNormal DistributionScientific Method, and Opportunity Cost. Of course there are many more models and some are job specific, such as financial models.

An abundance of people in the financial industry have models of how our financial markets work. Some models are simple mental models, others are complex Excel worksheets, and some are adaptive computer models, also known as machine learning or artificial intelligence. Innovation is inevitable, machine learning is coming into so many facets of our lives; therefore, financial advisors must learn how machine learning works and is used in the financial services industry.

A machine learning example

Computer programs rely on inputs to produce an output. For example, a simple program may look like this:

Inputs: X = 5, Y = 4
Model: X + Y
Output: 9

Machine learning works by providing the computer with a training set of data. A training set is a set of data with historical inputs and the corresponding historical outputs.

There are many different models of machine learning, including neural networks, deep learning, inductive logic (e.g. self-driving cars), and, the simplest, decision trees. We will go through an example of how a decision tree machine learning model works. For our example, we have a set of data on homes in San Francisco and New York. The goal of our model is to determine where the home is located based on the inputs of elevation, price per square foot, year built, number of bathrooms, number of bedrooms, price, and square footage.

Pairing each set of input variables with the actual location of the home gives us our training set. The computer analyzes this data, determining the best input variable by which to fork our data; if the elevation (input variable) is above X, then home is classified as San Francisco, if not home is classified as New York. The computer also determines the specific value (X) to use as a divider in the fork.

So, we have just forked our data into two different segments by elevation, homes on one side of the fork are classified as New York while the other side classified as San Francisco. However, by classifying the values at one singular point, we have misclassified many homes as San Francisco that are New York and vice-versa, as there is no singular elevation at which we can classify all homes above as San Francisco and all homes below as New York.

To rectify the misclassifications, we fork each classification AGAIN by another input variable. Sticking with our example, we take all the homes that were initially classified as San Francisco and fork it by price, with homes over a certain price classified as New York and those homes under a certain price classified as San Francisco. Some of the homes will remain classified as San Francisco, while others are reclassified as New York, giving us a more accurate model to use. Now, the computer applies this process to many subsequent layers in the decision tree, forking the data until it is 100 percent accurate on the training data.

After we have built our decision tree model, we can test it against real world data. Running inputs through the forks we created from our training set, we predict whether the home is in San Francisco or New York. The goal is to achieve perfect accuracy; however, using such a simple model, it is likely that we will not achieve perfection. While this is not ideal, it is a reality in certain models. As stated in the introduction, while all models are wrong, having a model with upwards of 95 percent accuracy can be extremely valuable when looking at large amounts of data.

How does this relate to financial services?

Understand that machine learning computers take large amounts of data and find patterns between the inputs and the outputs that humans would never see. The computer weights the corresponding inputs appropriately and performs advanced predictions on real world data. So how does this influence the financial advising space?

Robo-advisors use machine learning in many different ways. One way is to place clients into the correct portfolio. Fact finding sheets are filled out by the clients. The computer uses a decision tree model, like the one described in our example, to place the client into the appropriate portfolio based on the clients’ risk tolerance.

In fact, Intelligent Portfolios™ from Schwab offers a great view of how this works. Schwab requires the user to fill out a multiple choice fact finding sheet about risk tolerance. On the left, Schwab shows a portfolio allocation by asset class. Schwab displays how the asset classes change depending on the answer chosen. For example, answering yes to “Do you make financial decisions confidently?” will allocate more of your portfolio more into Stocks and Commodities (higher risk assets).

Robo-advisors also uses machine learning for portfolio rebalancing, portfolio tax minimization by asset class, portfolio tax-loss harvesting, and portfolio investment customization. As you may have noticed, every feature listed previously uses the adjective “portfolio.” While maximizing portfolio returns is one facet of investment advising, it is only one tree in the forest.

Being a financial advisor is now about being a life coach and guiding your clients through all the complexities and emotions that our financial world presents. Every advisor knows about how life stages effect a clients’ needs, desires, wants, and wishes. The competitive advantage advisors have is the comprehension of their clients’ emotions, wants, and needs – this is an extremely complex interaction unfit for robo-advisors.

We live in a world where people’s wants and needs are infinite, yet we only have finite resources to satisfy those needs, including the most valuable resource – time. People’s needs and wants are often not articulated clearly either – they are displayed with emotion absent of logic. Assessing a person’s goals, ambitions, passions, and personality is not a computer program. That is where financial advisors come into the equation.

Click here to learn more about the industry leading financial technology solutions offered by Advicent, or call (855) 885-7526 to speak with an Advicent representative.

Sign up for the
Advicent newsletter