Eli Butters: Machine Learning Algorithms in Quantitative Finance

My name is Eli Butters and I am a sophomore studying Statistics and Computer Science. I started out studying Finance and Real Estate as a freshman but then moved over to Statistics and Computer Science because I love understanding the mathematical side of finance. Over the past 20 years, virtually every market inefficiency has been arbitraged away and has left exceedingly random and competitive capital markets. With just our brain power, we can’t look at the market in any short term period and make any assumptions about underlying movements while producing any long term profit. However, Machine Learning Algorithms are specifically designed to recognize patterns in seemingly random data.

Eli Butters, Sophomore, Economics

Machine Learning (ML) Algorithms have been used in Quantitative Finance for much longer than most people think. The Efficient Market Hypothesis, created in 1965, led to the belief that markets were completely efficient and it was impossible to beat a passive investment with active investing. Although this is true for most active strategies, quantitative investors have been able to utilize ML models which allow non-linear relationships to be formed. This result has led to a stark growth in the use of these models throughout finance. 

ML models can be used creatively through any part of the process when investing. However, they are best suited for the cases of return forecasting, portfolio construction, and risk modeling. While each model can be used for a variety of different tasks, Multilayer Perceptrons (MLPs), Support Vector Machines (SVMs), and Long Short Term Memory (LSTMs) models are some of the most common choices for our three use cases. While MLPs and SVMs are extremely powerful, their ability to form efficient models given a large amount of features is far below that of an LSTM. For example, if someone were wanting to create their own quantitative investing portfolio, they might use an MLP or an SVM for the less intensive tasks of portfolio construction and risk modeling. One would be better off using an LSTM for the process of return forecasting because it involves so many features. 

Artificial Neural Networks (ANNs) work by taking input from a large training data set and then applying a backpropagation algorithm to the weights in the network through a cost and activation function. Recurrent Neural Networks (RNNs) are different in that they memorize past inputs by passing them through the network multiple times. You can think of this as passing the input data through multiple Feed-Forward Neural Networks. RNNs do this because of the underlying assumption that each data point is not independent from every other data point. Rather, each past data point depends on the prior element in the sequence. RNNs also differ from other ANNs when it comes to backpropagation. While RNNs still adjust their weights through a gradient descent approach, they use a backpropagation through time algorithm which allows them to sum errors through different data points, therefore remembering the effects of the previous data. 

We know that RNNs are better suited for time-series data because of their ability to remember previous inputs, although they break down when long term memory is needed for inputs. Because of their backpropagation through time learning algorithms, they weigh much more heavily towards recent training data. To fix this, Hochreiter and Schmidhuber created a Long Short Term Memory (LSTMs) model. LSTMs use Forget, Input, and Output gates within their neurons to logically remember or disregard certain data from its memory. This allows LSTMs to use relevant data from any time period in the past in order to predict the current input.

Throughout my summer, I will experiment with RNNs and LSTMs to learn how to better train them to predict more profitable trades. After proper data preprocessing and feature engineering, I will first test these networks with traditional market data such as open, close, and volume feature vectors. After I test the outcomes of these features, I will then move onto more complex features. These more complex feature vectors could prove to provide more alpha, a term used to signify a leg up on other market participants.

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s

%d bloggers like this: