We’d like to introduce you to Justin. A quant who has worked at places like Blackrock scientific active equity team. An elite team of data driven quants. Justin helped us develop the program and drew upon his background in natural language processing to design the first project in term two. So, Justin, how did you get into quantitating? I started off working for a consulting firm. I used some of the risk model tools like MSCI Barra to analyze client equity portfolios for style risks. From there, I went to Barclay’s Global Investors, a predecessor of Blackrock in their scientific active equity group. Could you describe the process that quants go through in generating Alpha? Do you start with the data? Do you start with the model? Do you start with an idea? Quants follow the scientific method. So, first we need to have a thorough understanding of finance, of how companies work and what can affect the company’s performance. From there, we form a hypothesis. Then we use tools and data to test that hypothesis. It’s important to start with the idea first instead of just jumping into the data, otherwise we’d be data mining. Patterns in data may form due to random chance or real underlying phenomenon, but strategies based on the former will not work well on unseen future data. We also read lots of economic, accounting literature, and attend financial seminars. So, you built one of the projects in term two. What are students going to do in the project? What were the main things you wanted them to learn? My main goal was for students to practice several fundamental natural language processing techniques and to see an example of a practical way that can be applied to make stock picks. The project is based on the stock selection model from an academic paper called lazy prices. Elements which I’ve used in practice. Students use NLP techniques to scrape text information from companies annual financial styling and analyze them to produce a signal that indicate whether the company’s stock price will appreciate or depreciate. That sounds really cool. It sounds like one example of a way new techniques are being used in quantitative finance. How else are AI and machine learning being applied in quantitative finance? AI and machine learning models are tools that quants can use to enhance their existing research method. There’s a lot of alternative data such as credit card transactions, satellite images, and environmental social governance data. I’d say a major source to opportunity and active application in finance is with text data. Whether it’s news, social media, financial reports and earnings calls. The reason is that ideal data set is one that can have broad coverage of many stocks. Price volume, and fundamental data fit this criteria and so does texts. So, natural language processing is one of the main applications in research and strategy phase of investment, where quants process data to generate Alpha signals that can be fed into portfolio optimization. For example, quants have known and tried various ways to measure attentional bias. That is when certain stocks are mentioned more frequently in the news. This tends to increase trading activity and volatility for those stocks. One way to measure this is by calculating the skewness of volatility on price volume data. Another way is to directly measure the level of news reporting on each stock using natural language processing. There’s also research into processing satellite images to generate signals. I would point out that for Alpha signals to be most useful, they would ideally generate signals for a broad set of companies. This is still a challenge when it comes to satellite data. For example, there’s been work done to track parking lots and estimate the performance of retail companies. But an ideal Alpha signal is one that would apply broadly across sectors and industries. Another part of the quant workflow where AI is being used is to enhance the alpha combination stage. Traditionally, quants domain knowledge and research to propose combinations between signals. You may hear of these as conditional factors and we often refer to these as interactions. Quants can also train a model to find these complex interactions among factors. There’s also been research into applying reinforcement learning to finance. I should point out that at the moment, reinforcement learning is showing more promise in trade execution not an Alpha generation. It sounds like there’s a lot of opportunity to utilize new techniques and analyze new data sources. What are some of the challenges of applying AI to quantitative investment? Well, as I mentioned before, AI and machine learning are just some of the many tools available to understand and predict company performance. In some ways, when you have a hammer, everything looks like a nail. Well-established economic methods are still the best tools for certain tasks. There are definitely some important challenges to consider when applying AI techniques and trading. In finance, the signal-to-noise ratio is incredibly low. Financial data have relatively little depth and time. Some data, especially newer alternative data, just have not been collected for very long. So only a few years worth of data exist. There are challenges unique to time series analysis and that there’s only ever one instantiation of time. There’s directionality of time. Financial time series are non-stationary. It’s very easy to produce models that are over fit to the training data. So, it’s critically important to take steps to avoid over-fitting. Since there’s such a huge risk of over-fitting, a key challenge is to develop models that are as simple as possible but not simpler. Finally, it’s typically very important to be able to explain why a strategy works to managers investors. Investors may not be comfortable investing in a black box strategy. For this reason, being able to interpret models is very important and it rains a limitation of many AI tools. With all the advances seen in AI and machine learning, where do humans fit into all of this? Is there’s still a need for human intuition? Yes. We still need human intuition and domain expertise to do a sanity check on our models. For instance, your team may have several models such as outputting signals for every stock. It’s good practice for quants to quality check a handful of the outputs. For instance, we typically look at the top 15 and bottom 15 stocks that are ranked by the model and make sure that the model’s output makes sense. Some checks may include watching out for stocks where there may be a bio target or that may be exposed to future litigation. Remember that we use AI and models when it’s possible to automate some of the data processing. But there’s always going to be information that can’t be distilled and fed directly into the model. That’s why it’s important for quants to have a good understanding of how financial markets work and still follow financial news very closely. Also, a model is only as good as the data is fed into it. A great deal of work goes into the quality assurance of the data. In fact, there are entire teams that are dedicated to quality checking the data on a daily basis, because this is so critical to a firm’s performance. In order for a quant to effectively apply AI to generate Alpha, it’s also important for them to have a sound understanding of the data of the model. At least for the foreseeable future, human intuition is critical to the success of an AI based investment strategy. Well, this has really been great. Thank you so much for your time. One last question. What advice would you give to someone just entering the field? I would say, try to develop a broad understanding of as many different models and tools available as you can. Deepen your understanding of the application of the models by developing personal projects and be able to communicate the steps you took when developing. A key part of our role as quants is not only in building the strategies, but in understanding them and communicating how they work.