BG Interview: Machine Learning in Investment Management

January 18, 2019

Listen to the interview here

Hello welcome to our 4th Quarter Interview. I am Jamie Houston of the Bristol Gate Relationship Management team. This quarter instead of speaking with our Portfolio Managers we have Leyla Imanirad, Senior Research Associate on our Data Science team.

Leyla has been with Bristol Gate since 2011 and been instrumental in designing and implementing our machine learning model which we use to predict dividend growth.

Over the last year, Leyla has represented Bristol Gate at a number of conferences and presented across Canada and the United States about Artificial Intelligence, Machine Learning and how Bristol Gate is at the forefront of integrating it in Investment Management. Today Leyla is here to help answer some common questions we have received from investors.

Leyla, thanks for joining us.

Jamie Houston: So Leyla, we hear more and more about the importance of Artificial Intelligence. Why now?

Leyla Imanirad: It’s a very good question considering the field is more than 60 years old. The term, Artificial Intelligence, was coined in the fifties with contributions from scientists with diverse backgrounds such as neuroscience, philosophy, mathematics, and computers. These scientists got together in Dartmouth and built a common vision: designing systems with human intelligence behavior. Given the state of computing at the time, let’s just say, this was an ambitious dream. As humans, we can do pretty dumb things but we’re still smart enough to dream of and try to replicate Human intelligence behavior.

Now, what is the most markedly characteristic of intelligence, shared by all living creatures? The ability to learn and to adapt which is the basis of Machine learning! We define Machine learning as a collection of algorithms that give computers the ability to learn from data. What do we mean by learning? Think about a little kid learning about animals for the first time. As you show them pictures or videos of cats and dogs and dinosaurs, they eventually learn to distinguish them from one another…..  In our brains that almost happens automatically, without much effort. In computers, we mimic that with sophisticated ML algorithms.

Why now?

But why are we experiencing the excitement today even though the core of AI and ML was developed many years ago?

1. infrastructure: We finally have the software & hardware infrastructure necessary to make the technology accessible to a large group of people. We have open source libraries implementing different machine learning algorithms, we can rent a GPU in the cloud, which makes “AI as a service” a commodity rather than a novelty.

2. data: Artificial intelligence is data hungry! You need data to build these models, the more the merrier. And we live in a data-rich world and it is believed that the amount of data we generate doubles every year! Think about everything we can learn from the 500M tweets per day!

3. Investment: the governments, academic institutions, large enterprises, and even small companies all want to leverage the benefits of the technology. People are making investments and it pays off.

JH: Very interesting. Bristol Gate has been using technology to predict dividend growth since its founding in 2006. Can you take us through how our model has evolved and how our latest version predicts dividend growth?

LI: The idea of predicting dividend growth was one of the main pillars of our investment philosophy from the early days. Back in 2006, our CIO and co-founder of the firm, Peter Simmie, developed the first version of our model which was a regression model to predict dividend growth 12-months forward. It was an excel-based model using annual fundamental data. This model was later implemented & refined in SAS, a statistical software.

A few years later (2014), we upgraded that model to a two stage model, consisting of a classification and a prediction step. In this model, we would first classify stocks into different buckets based on their dividend growth behavior and run a different regression model for each category.

As we got access to better tools and better data, we moved from SAS to python and developed a ML-based algorithm to predict dividend growth for the stocks in our universe.

With each iteration, we would test the model for prediction accuracy and consistency. The objective is to have good predictions for many stocks, with as few bad predictions as possible. We would also usually run the models in parallel for some time before adopting the new model in our investment process.

Model construction:

Every modeling process starts with data & we pull about 20 years of history for each stock in the US universe. The feature set that we build consists of fundamental data such as dividend, and cash flow growth, as well as historical price and estimate data as available. The data set that we build will have more than half a million records, one row for each stock at a given point in time, and more than a 1000 features as columns. We use this data set to train GBM, which is a tree-based model. The algorithm decides on the features that are most relevant in dividend growth prediction, reducing human biases in the modeling process. Once the model is trained on historical data, we can then build a data set from the new data as it becomes available and apply the model to come up with the predictions for next year.

JH: We get this question quite often. When people hear “model” in investing they automatically think a Quant strategy. How does Bristol Gate’s process differ from a purely Quantitative strategy?

Quant model:

LI: That’s a tricky one to answer. AI has been used in quant strategies before but mostly for short-term prediction & trading, however that trend is changing now. I would say we’re one of the few investment managers using ML in predicting fundamentals over a relatively long time horizon of a year.

I believe there is another big difference between our strategy and a traditional quant strategy:

In many quant (or smart-beta) strategies, there is very little human involvement. As a portfolio manager, you decide on a factor driving a strategy and the machine makes the decisions based on each stock’s factor exposure. For us, predicting dividend growth is a first step in our man & machine approach, meaning, we simply use this step to build a focus list of top 50-60 securities that we want our fundamental team to analyze. This allows them to focus their efforts on a smaller set of securities, saving time in the analysis step.

JH: Thanks Leyla. We hope that you have enjoyed the interview. We now ask that if you have any questions, please do not hesitate to send them in and we will do another interview to answer them.

For those of you looking for detailed information on 4th quarter performance it can be found in the Investment Letter at BristolGate.com under Insights.

Thanks for listening.