Erez Katz, CEO and Co-founder of Lucena Research
At Lucena we look to turn knowledge into actionable insights. There has been a lot of buzz surrounding machine learning for Finance to forecast securities. We wanted to host a webinar to show a unique approach we are utilizing to identify investment opportunities through deep neural networks. More specifically, how convolutional neural networks (CNN) provide a compelling approach to classifying timeseries data in order to project stocks’ impending price action.
What you can expect from the webinar:
- A quick explanation of what deep neural networks are and how convolutional neural networks (CNNs) classify images.
- How CNNs are used to recognize hand writing with uncanny accuracy (above 99.7%).
- Discussion on how the success of computer vision using CNNs for image classification, speech recognition, and object detection can also be applied to tradeable securities.
- How our trial and error led to a compelling solution. Specifically, we will spell out what seemed promising on paper, but didn’t work for us.
- Lastly, we will describe what specific actions we took to “help” the neural networks learn and how we attempted to overcome data not always conforming to IID (Independently and Identically Distributed data) and the non-stationary nature of the financial markets.
Whether you’re an investment professional looking to understand machine learning or a Quant with experience in quantitative finance this discussion has something for you.
It’s important to distinguish between two forecasting approaches:
- Forecasting the future price of a stock within a predetermined timeframe. For example: Apple will reach $200/per share by August 30, 2018
- Classifying if a stock is destined to move in a meaningful way relative to a predetermined benchmark. For example: Apple will move higher relative to the S&P by at least 1% within the next 21 trading days.
Market-relative classification (Option 2) is easier to tackle since outcomes are measured as Booleans (true or false). In addition, identifying a price move relative to a benchmark isolates our analysis on the idiosyncratic merit of the security while removing dependencies on its sector, industry, or the market it is traded in.
For example, assuming the market drops by 2% in the next 21 days and Apple (APPL) only drops by 1%, our classification would still be considered successful. Conversely, if the market rises by 1% in the next 21 days, we would expect APPL to rise by at least 2%. In general, best practice guidelines assert that if you can consistently classify benchmark relative returns with 53% accuracy or higher, you have enough statistical significance to overcome slippage and transactions cost and consequently you are well on your way to construct a compelling investment strategy.
In the Webinar, we will showcase an example in which we were able to achieve 69% accuracy out of sample. While we stand firm on the authenticity of the results, we are not claiming that it’s enough to put forth an algorithm that will always beat the market. Those of you managing money professionally know well that the dynamic nature of the market and the speed in which an investment opportunity gets exploited to obsolescence are factors that should be considered at all times.
Our goal is to give you a glimpse into the considerations that a quant team takes into account while attempting to provide enough actionable reference for the portfolio managers to succeed.
Here’s the full list of Q&A we received during the webinar. Enjoy!
Want to learn more about our flagship machine learning platform QuantDesk? Here’s an overview. Still have questions? Let’s chat.
Liked this post? Read more about similar topics: