In this module, we'll look at the evolution of data-driven models and their applications in trading and finance. First, we look at the three main approaches to modeling financial markets based on their data and behaviors. Then we'll explore the use cases for machine learning and trading investment and finance. We'll then focus on the applications that have the highest potential to benefit from the application of machine learning and AI. We'll also discuss the challenge of creating an interpretable ML model. Models guide most trading and investment decisions which are essentially predictions of relative performance. They're created by training an algorithm on a training data set and tested by applying the model to a testing data set. Models can be distinguished by the type of algorithm used. Traditional models of financial markets use purely statistical algorithms such as regression to make predictions. To improve the predictive power of models, machine learning and reinforcement learning algorithms are employed. Earlier models view the market's pure data generating processes similar to the approach taken in econometrics. They apply statistical algorithms to market data to reveal the essential nature of this process and then make inferences and predictions to guide trading and investment decisions. These models are relatively easy to interpret but may make inaccurate predictions. As the market's essential nature is assumed to be relatively static, the model might be tweaked but it's never overhauled unless it fails dramatically. Machine learning models view the markets as an ever-changing collection of behaviors. Teachers and loadings need to be updated continually in the service of maximizing the model's predictive power. Models created with ML Algorithms are more complex and so are difficult to interpret statistically. This loss of interpretability is offset by their potential to make better predictions and more quickly adapt to structural changes in market behaviors. Models created with reinforcement learning algorithms attempt to create an autonomous intelligent agent that learns by experiencing new market states, taking actions, and making gains or suffering losses as a result of those actions. The agent is endowed with a memory and to adjust its strategy to maximize profit. Reinforcement learning models seem simple, but they lack interpretability due to the complex feedback loops in their long short-term memory networks. Machine learning algorithms have been used successfully to create financial application for some time now. Given the high volume, accurate historical records, and quantitative nature of the finance world, few industries are better suited for artificial intelligence generally and machine learning in particular. There are more use cases of machine learning and finance than ever before, a trend driven by more accessible computer power and more accessible machine learning tools such as Google's TensorFlow and BigQuery ML. We'll explore the use of machine learning techniques and algorithmic trading, portfolio management, credit decisions, and fraud detection. Machine learning is a powerful technology which helps to analyze thousands of data points within seconds. It can identify emerging price patterns rapidly, compare them to historical experience, and initiate a trade when the odds of success are favorable. An AI that combines an analysis of sentiment and news headlines and social media comments with market data from multiple sources is able to predict both the direction and magnitude of moves and asset prices. ML also facilitates trading every millisecond with no human intervention needed or even physically possible. With this origins going back to 1980s, algo trading involves the use of complex AI systems to make extremely fast trading decisions. Algo systems often make thousands of trades in a day, hence the term high-frequency trading. Hedge funds and other trading firms do not openly disclose their AI approaches to trading. But there are clear indications that machine learning and deep learning are playing an increasingly important role in calibrating decisions in real time. Machine learning algorithms can analyze hundreds of data sources simultaneously, something that human traders cannot possibly achieve. They help human traders squeeze a slim advantage over the market average. And given the the high volume of transactions, that small advantage can translate into significant profits. The term robo advisor was essentially unheard of six years ago but it's now an established feature of the fintech landscape. Robo advisors such as Wealthfront and Betterment are algorithms built to calibrate a financial portfolio to the goals and risk tolerance of a specific investor. Investors create profiles that contain their goals, for example, retired at age 40 with a million dollars and their age income and current financial assets. The advisor, or more accurately the allocator, then spreads investments across asset classes and financial instruments in order to reach the investor's goals. The algorithm also calibrates to the changes in the investor's goals and to real-time changes in the market, aiming always to find the best portfolio allocation. Robo advisors have gained significant traction with millennial consumers who don't need a physical advisor to feel comfortable investing and who are less able to rationalize the fees paid to human advisors. At big banks, machine learning algorithms can be trained on millions of examples of computer data such as age, job, income, and financial lending results such as whether or not a person defaulted or paid back their loans on time. The underlying trends in these factors can be assessed with algorithms and analyzed continuously to detect trends that might impact lending such as they're increasing rates of default among a specific demographic over the last two months. These results were extremely valuable for lenders as they can cut their loan losses by anywhere from 6 to 15%. But at present time, they are used primarily by larger banks that have the resources to hire data scientists and with access to the massive volumes of past and present data to train their algorithms. Security threats and finance are increasing along with a growing number of transactions, users, and third-party integrations. For instance, banks can use a machine learning algorithm to monitor thousands of transaction parameters for every account in real time. The algorithm examines each action a cardholder takes and assesses if an attempted activity is characteristic of that particular user. Such models spot fraudulent behavior with high precision. If the system identifies suspicious account behavior, they can request additional identification from the user to validate the transaction. Or even block the transaction altogether if there's at least 95% probability of it being a fraud. Machine learning algorithms just need a few seconds or split seconds to assess a transaction. The speed help prevent frauds in real time, not to spot them after the crime has already been committed. Adyen, Payoneer, PayPal, Stripe and Skrill are some of the notable fintech companies that invest heavily in security machine learning. Now we'll take a look at three use cases where machine learning is likely to expand its footprint. Applications of automated financial product sales existed [INAUDIBLE] some of which may involve machine learning but others are just rule-based systems. A robo advisor might suggest portfolio changes. And there are plenty of insurance recommendation sites that might use some degree of AI to suggest a particular car or home insurance plan. In the future, increasingly personalized and calibrated apps and personal assistance may be perceived as more trustworthy and objective and reliable than in-person advisors. Just as Amazon and Netflix can recommend books and movies better than any human expert, ongoing conversations with financial personal assistants might do the same for financial products as we see already happening in the insurance industry. Much of the future applications of machine learning will be in understanding social media, new trends, and other data source, not just stock prices. The market moves in response to many human-related factors that have nothing to do with prices or trading volume. The hope is that AI and machine learning will be able to replicate and enhance human intuition of financial activity by discovering new trends and telling signals. Data security risks thrives in an environment where the Internet is used for most transactions and where an increasing amount of sensitive company data is stored online. Machine learning algorithms can significantly enhance network security too. Data scientists train assistants to spot and isolate cyber threats as machine learning is unrivaled in analyzing thousands of parameters in real time. And chances are, this technology will power the most advanced cyber security networks in the near future. There are some downsides to machine learning and deep learning methods. One drawback of ML is that you may struggle to explain why a model is doing what it does. This is commonly known as the black box criticism. The process by which deep learning techniques reach decisions is also unclear. Deep learning techniques provide predictions, but do not provide insight into how the variables are being used to reach those predictions. This is especially important for trying to prevent discrimination in lending and other consumer models. The lack of interpretability can be problematic in use cases ranging from medical interventions to autonomous trading, criminal justice, risk management, and many other areas of society. In many cases, the usefulness and fairness of these AI systems is limited by our ability to understand, explain, and control them. Google Clouds Advanced Solutions Lab has invested considerable effort and research in to unlocking the black box of powerful yet complex ML models like deep neural networks. Explainable AI refers to the collection of methods and techniques which enable a human to understand why a model is giving specific results. You can find a complete description of explainable AI by Chris Rawles along with the code he used in the two references linked below.
