Demystifying the Business of AI, Machine Learning, and Big Data

The tech in the investech sector really comes in two flavors. There are a lot of new technologies that make the investment process more efficient and more accessible. We've seen some of that in the first module of this course. Automation in asset allocation, financial planning, and security settlement have really started to drive down cost across many investment sectors and enable new developments like Robo advisers, automated financial planners and smart index funds to make the average investors' portfolios more customized and more diversified and at the same time less costly. The other category are the technologies that make the investment process more precise and these are broadly labeled sometimes incorrectly as AI, big data and machine learning technologies, which are simply data analytics tools that enhance our ability to interpret the historical financial data and make more accurate predictions for the future. We've talked about some of these such as neural networks primarily from a technical point of view in this. Module will take a broader business focus point of view on the application of AI and machine learning techniques in the investment industry. Will take a primarily non-technical approach in this module with the goal of providing an industry overview did mystifying the key buzzwords and clarifying some common misconceptions providing some practical advice and directing you to the more in-depth research and analysis in the relevant subject fields such as computer science.

In statistics, should you be interested in pursuing further? First let's take a look at what Ai and or machine learning really are as these concepts are probably on par with crypto as some of the most over hyped terms in finance. First of all, there are two types of AI systems one is called narrow AI which is defined by divorce key is a computer system that matches or exceeds human intelligence, but in a narrowly defined area under this definition most computer programs can be called narrow AI to some degree as most of them would exceed human's capabilities in some area next. The general AI is probably the one that you're more familiar with and that's essentially The Skynet is a software. But without the pre-program rules an artificial neural network that is truly able to code itself and rewire self with new data patterns. Essentially the general AI is a true imitation of the human brain so far, the current computing capabilities of our hardware is far from being able to even remotely construct a general AI network. So most of the research in this type has been theoretical therefore the current application of all of the so-called AI systems in finance has been at best narrow AI, a more appropriate label for AI in finance though is probably machine learning you see the financial industry particularly. The investment sector has been all about processing data and information extracting new insights and signals from current and historical data. In these settings the use of AI has essentially been the use of more and more advanced models of machine learning.

Which is self is a fancy buzzword for statistics essentially using machine learning or quote unquote AI investments is using advanced statistical models on the same financial data to better interpret patterns in the data conduct better statistical inferences or make better predictions. Now, let's put the tools and the data together. To have a more complete picture of the type of the data out there in the investment landscape and the type of tools that you can use on them first. There are all kinds of data that you can use to guide your investment decisions out there numbers, reports, text, news, you name it. Now, these data will belong to one of two types. Let's call the first type quote-unquote small data. These are the data small enough in size like files and excel spreadsheets that you can store. Or on your computer and search on your computer and the counterpart to that is big data. And what's Big Data? Well data set that's large enough that you can't store it on your own computer, but instead have to put it in a dedicated storage server and have to use some specialized database tools like Hadoop and Mapreduce to efficiently process and manage the data. That's it. That's literally the bus free meaning of big data now within both the small data. And dictate the categories data can further be classified into two types based on their structure. The first type is called structured data. These data are essentially numbers, numbers that you can organize neatly in rows and columns like in an excel spreadsheet now surprisingly. It's a little easier to perform analysis on structured data because there are already quite organized in contrast to that are unstructured data. These are the natural languages, text, images, videos, etc. Because these data are not neatly stored as numbers in rows and columns that will take a little bit more effort to analyze first. We'll have to use some statistical technique like natural language processing or NLP for text data and signal analysis for audio and video data to convert these data into numbers before we can analyze them but the important thing to note here, is that contrary to what you hear in the hype. There's no mutual exclusivity between the data types. For example, not all unstructured data are big data things like text or image files could be very small in size. So you wouldn't need mapreduce to work with them, but they still require the extra processing step to be converted into numbers.

There could be small unstructured data and there could be big structured data, now with all that data available. Here are all the tools that you can use to mine these data for insights and signals starting with the simplest. We have what I call the simple analytics like basic summary statistics, for example, averages standard deviations correlations etc. Now basically everything else can be labeled as machine learning which again is just a fancy buzzword for the more advanced statistical models beyond simple statistics, like averages that you can use on the data to analyze patterns conduct inferences or make predictions next. There are two types of machine learning tools the easier type. Let's call that quote unquote shallow learning a classic example of shallow machine learning is a linear regression, shallow learning is essentially a statistical model where we the user have to specify most of the model parameters. For example, when you're estimating a regression, you have to choose what are the independent variables the regressors that you want to include in that regression. The counterpart to that is deep learning and the bus free meaning of deep learning is essentially a statistical model with more parameters that can be directly determined by the data an example of deep learning is regularized linear regressions, like the lasso or the elastic net. These are the regression models where instead of you choosing the independent variables you low the entire variable list to the model and based on some basic criteria. The model will choose the most quoted. Quote important regressors to include for you that's essentially deep learning, letting the data determine the model parameters from the last module. We already saw that neural networks or another classic example of deep learning. And this class of models is the closest to quote unquote and AI that we can use for financial data analytics and further within each machine learning type. We also have two sub categories. The first category is called supervised learning and as its name suggests supervised learning models have to be trained with data before we can use them for tasks like predictions again, a linear regression is a classic example of supervised machine learning. Or lending regression examples from the credit tech class any regression model there has to be fitted with data first. We need to get the parameter estimates for these alphas and betas and only then can we plug in new data and get the predicted values like the default rates by this token, you can view other much more advanced supervised learning models as super souped-up regressions. The other class of machine learning models are unsupervised. This means that we don't need to use existing data to estimate the model before we can use it. We can just directly use it a classic example of unsupervised learning models is clustering given a bunch of data points. You can directly take them to a clustering model and the model will cluster them into groups where points within a group are close enough. We've seen a model like that in action in the platform landing module of the credit tech course, so basically both shallow learning and deep learning can either be supervised or unsupervised finally to put the data and the analytics together. Well, you have complete freedom to choose whatever tool you see fit to analyze the type of data at hand. You can use a supervised shallow learning tool like a regression to analyze structured big data and there are unsupervised deep learning tools for unstructured small data when it comes to using AI or machine learning for investments. You're really limited by number one. Your computing power and number two choosing the most appropriate model that works best for your data type and your questions.

Let's now take a deeper look at the usage of unstructured data in investments because with the competition for Alpha is getting more and more intense. Investors are increasingly turning to the analysis of unstructured financial data under using ever more sophisticated models to mine them, if you think about it in finance the bulk of the growth in data that we have available has been an unstructured data. This is from Apple's SEC 10K filing is annual report for 2019. And this income statement here is an example of structured data with numbers neatly stored in columns for tabulation and analysis. And when released these data are obviously the first focus of analysts and traders alike. However, think about how long this annual statement is the tables are only a few pages at best and the rest of the statement that hundreds of pages of text is all unstructured data. Much fewer people pay attention to these not to mention finishing them as they're not very exciting reading but these languages might contain a lot of information. What worse and what sentences are used and how they're put together can tell you a lot about what the management might be thinking beyond just the income numbers, same for the social media data. This is a random Twitter page with posts about Apple stocks at the end of 2019. As you can see these social media languages how positive or uncertain they are for example can also tell us a lot about investor sentiment and psychology around the stock. Of course, there's way too much of these unstructured data text images that are relevant to the stock out there for us to read them all and that's precisely where computers can help. For example, we can use natural processing tools to first convert the text into some form of numbers that we can fit our existing machine learning models on at the simplest level. For example, the whole 10 K filing could be turned into a word vector with entries denoting to the frequencies of each English word that's used in the document. We can then stack different document vectors together into a matrix and run all sorts of shallow and deep machine learning algorithms on it to get an estimate for both content and on text like how positive the languages are how complex the sentence structure is how many topics are talked about among many other things. So I encourage you to explore a bit more about natural language processing and other methods for unstructured data analysis, because when used appropriately they can really extract information and signals from these data that are truly incremental to what you get from purely the numbers.