[Caleb White is a first year MBA student at the Nicholas Center for Corporate Finance and Investment Banking at the Wisconsin School of Business. Caleb utilized one of his electives to take a new course called Machine Learning for Business Analytics]
Big data is a trendy topic in the business world. Analytics and machine learning are frequently popping up in job descriptions. Excel is far and away the leading tool for business professionals, but what happens when you have too much data or too complex of problems for Excel to handle? When Excel isn’t enough, what tools are available and what do they look like? What is machine learning and why should we care about it?
When I considered options for a full-time MBA program, my desire to learn more about data mining and analytics was something that I found of mutual interest in the Nicholas Center. This Spring I was given the opportunity to take a machine learning course as one of my electives in the WSB. I knew that I wouldn’t become an expert data scientist from one class, but I recognized that data has a major role in the future of business, and I wanted to see how it could impact a finance team. After the first 8 weeks in the course, I am excited about what I have learned thus far and want to share my experience.
Machine Learning
In basic terms, machine learning refers to the process of detecting meaningful patterns in large data sets and developing a tool belt for understanding what the data is saying. Detecting the patterns in large data involve selecting the right model for the data (think of it as using a hammer for nails or a wrench for bolts), and building an algorithm that balances complexity and adaptivity (think of the algorithm as adjusting the wrench to fit the current bolt and hopefully any other bolts you come across).
The goal of machine learning is to focus on inference from data to build a model for predicting the future. The key of determining a strong model is the bias-variance tradeoff. Statisticians for many years viewed “bias” as an error that could easily be removed and focused on finding the best unbiased estimators. However, machine learning plays with the idea that bias is not always a big thing because a small amount of bias could lead to a massive reduction in variance leading to a model that does a better job predicting the future (i.e. performing better on out-of-sample data).
Model Selection
The most important step in any process is selecting the right tool for the job. Many people have heard of regression analysis or even seen linear regressions run in Excel. These often involve an x-y plane with plenty of data points and a trendline fit somewhere in the middle to provide a modest estimation of the data and its variables. Understanding linear regressions1 is the first step in machine learning, but the models quickly accelerate into classifications2, ridge and LASSO regressions, splines, discriminates3,4, and neural networks.
While these models vary in coding and mathematical complexity, they all aim to do the same thing: Use the data we have to predict the future. A rule of thumb from class describes selecting the correct model as, “We want the model to be as simple as possible but as complex as necessary”. The most complex model is often not the best model. Not only will it likely fit the data poorly, but complexity breeds risk of human error in the coding and difficulty interpreting the results. When done correctly, a model transforms data into a simple, compelling message.
Coding Programs
Coding is the reason many shy away from learning more about data analytics. I had almost no knowledge of coding before taking the class but took this opportunity to dive in head first. I feel that the codes behind machine learning are just like knowing any other language of business. There is no need to be an expert, but having a base knowledge goes a long way to understanding and discussing specific algorithms and their results. What I have found is that the most-used analytic programs break out in the following way, from easiest to hardest to use:
- Microsoft Excel
- R
- Python
For those who are experts in Excel and use it daily, picking up R and Python are quite similar and easy to pick up. Formulas such as norm.dist and IF flow nearly identically to formulas in R. The catch is getting familiar with the interface and formatting. R has a steep learning curve but is intuitive and easy for quick analysis. R also provides helpful tools for any formula by pressing F1 within the formula, and there are hundreds of online resources that provide guidance and codes you can paste right into your own program. Python looks and feels cleaner but requires a deep understanding to be used effectively. Dedicating an hour or two to online tutorials and experimenting on small data sets provide quick and easy learning opportunities.
Why It Matters
Attempting to predict the future is prevalent in most business units, and many companies have plenty of useful data. These are prime areas to explore with machine learning. In a simple insurance claim data set we pulled, a LASSO model in R produced a mean-squared error (MSE) 20% smaller than a traditional Excel linear regression. Imagine the impact on a business if you could predict the return on advertising expense with 20% greater accuracy or remove 20% of the loss on an insurance portfolio. This was done with 180,000 rows of data for each of 130 variables. Excel churned the data for a few minutes, but R handled it without missing a beat. The ability of machine learning is astonishing.
There is a misconception that machine learning is too narrow or specialized to fit into most business models. The reality is that machine learning is so broad that it can be used in nearly every situation if we simply know how to use it. We are building a neural network model in April that uses wine tastings to predict sales of a new vineyard. I doubt the vineyard family ever considered using coding to predict the success of their next wine, but it’s clear that machine learning is here to stay and could play a major role in business for a long time to come.
Clockwise from top-left: Linear Regression, Nearest Neighbor Classification, Quadratic Discriminant Analysis, Linear Discriminant Analysis. Figures 2.1, 2.3, 4.5, and 4.6 of Hastie, Trevor, et al. Elements of Statistical Learning. 2nd ed., Springer.
Categories: