Rick Reviews/Rick’s Picks on AI/ML & Alt-Data

“Machine Learning For Trading”. Gordon Ritter, PhD. (2018). Gordon Ritter earned a PhD-Mathematical Physics from Harvard University (2007). Gordon is an Adjunct Professor, Founder/CIO at Ritter Alpha, LLC and was named “Buy-Side Quant of the Year-2019” by Risk.net. In Machine Learning For Trading, Dr. Ritter applies a “Reinforcement Learning” algorithm to generate trading strategies, size positions and minimize market impact on trading decisions. Market Impact –the tendency for large trades to push prices the wrong way – has long been a HUGE obstacle to implementing quantitative or discretionary manager trades.

Reinforcement Learning (RL) is the system that the Google AlphaZero/AlphaGo Zero that taught itself from scratch how to master the games of chess, shogi (Japanese chess), and GO, beating world-champions in each case. RL algorithms play different versions of itself (millions of simulations to master games) to get incrementally better by trial and error, learning by mistakes, through reinforcement. In reinforcement learning, software agents learn how to choose actions that optimize cumulative “rewards” over a multi-time period. In games of skill, the reward is winning the game by beating world champion chess or Go players.

Gordon Ritter’s reinforcement learning algorithm was used to discover and implement dynamic trading strategies to mitigate portfolio transaction costs. Ritter’s RL technique trained a computer to trawl through large datasets, simulate market impact and develop optimal trading strategy. It’s a unique computational solution to a “trader’s dilemma” of how to overcome trading and transaction costs that can eat up any available alpha. His research and application of RL to trading provides a “proof of concept” for superior portfolio optimization.

This reviewer was led Gordon Ritter’s research when searching for a portfolio optimizer based on a Kelly Bet Sizing/Kelly Criterion. {The Kelly Criterion or Kelly Bet size is a mathematical formula that helps investors and gamblers calculate what percentage of their money they should allocate to each investment or bet. John Larry Kelly was a Bell Labs scientist whose formula has become a part of mainstream investment theory.}

“Machine Learning for Stock Selection”. Keywan C. Rasekhschaffe & Bob Jones. Keywan C. Rasekhschaffe is a Portfolio Manager (PM) at Quantbot Technologies, LP in New York and has a PhD in Finance from the University of Lugano. Robert C. (Bob) Jones is the CIO/Co-Founder of System Two Advisors. In 1989, Bob Jones founded the Quantitative Equity Group at Goldman Sachs Asset Mgmt. These two portfolio managers turned to machine learning due to the low signal-to-noise in forecasting returns and equity returns are non-linear. Most securities’ markets do NOT follow the normal or Gaussian (bell curve) distributions. Both authors are practicing portfolio managers applying research to investment management.  

There is no one Machine Learning (ML) algorithm that gives the best results when applied to the different types of market data. Rasekhschaffe and Jones’ approach to stock prediction and selection involves “Ensemble Learning”. In statistics and machine learning, ensemble methods use multiple learning algorithms to obtain better predictive performance than could be obtained from any one machine learner algorithm alone. Ensemble methods combine several base models in order to produce one optimal predictive model. There is widespread evidence and documentation that multiple forecasts lead to increased accuracy.

Rasekhschaffe and Jones used a combination of four different machine learning algorithms:

• AdaBoost (Adaptive Boosting). A machine learning meta-algorithm that works on the principle of boosting. The core principle of AdaBoost is to fit a sequence of weak learners (i.e., models only slightly better than random guessing, such as small decision trees) on repeatedly modified versions of the data. The predictions from all of them are then combined through a weighted majority vote (or sum) to produce the final prediction. 
• Gradient Boosted Regression Tree. Decision trees build upon iteratively asking questions to partition data. Easier to conceptualize w/visual representation of a decision tree:

One decision tree is prone to overfitting. To reduce risk of overfitting, models that combine many decision trees are preferred. Each tree attempts to minimize the errors of previous tree. The combined models have better performance in terms of accuracy.

• Neural Network/Multi-Layer Perceptron (MLP). Multilayer perceptron is a feedforward artificial neural network (ANN) that generates a set of outputs from a set of inputs. MLPs commonly used to solve simple regression problems. Some MLPs have a single “hidden layer” (a vanilla neural network) – but these practitioners used four layers in theirs.
• Support Vector Machine (SVM). Support-vector machines are supervised learning models that analyze data for classification and regression analysis. SVMs are extremely popular because of their ease of use, calibration and ability to handle multiple variables.

Rasekhschaffe and Jones built 12 models – using three training windows for each of the four algorithms. They used a Big Dataset with an average of 5,907 stocks per month, covering 22 developed markets and assessed a total of 194 factors (firm characteristics). They divided the stocks into deciles based on output signal and created a long/short portfolio. They found that the ensemble machine learning model uncovered complex, non-linear patterns that are hard to tease out with traditional statistical methods. They also reduced the risk of model overfitting through feature engineering and combining forecasts in their ensemble model.

This reviewer was led to “Machine Learning in Stock Selection” while researching “ensemble machine learning models” for stock and stock index prediction. Ensemble machine learning  models extract information from various features and can exploit time-varying relationships between factors and return that aren’t captured by linear risk and alpha-generation models. Ensemble Learning has the potential to generate superior forecasts of stock returns in lieu of using a silver bullet approach of employing a solo machine learner algorithm.

Microsoft Azure. Machine Learning Algorithm Cheat Sheet – 1 Page Taxonomy

With over 100 tried and tested machine learning algorithms, there’s no one “silver bullet” algorithm to use in quantitative investment management. Here’s a Machine Learning Algorithm Cheat Sheet from Microsoft Azure to classify the type of tasks ML Algos can perform:

• Regression. Makes forecasts by estimating the relationship between values. Answers questions like: How much or how many?
• Two -Class Classification. Answers simple two-choice questions, like yes or no, true or false. A binary classification algorithm if you will.
• Multiclassification System. Answers complex questions with multiple possible answers. Answers questions like: Is this A or B or C or D?
• Clustering. Separates similar data points into intuitive groups. Answers questions like: How is this organized?
• Text/Sentiment Analysis. Derives high-quality information from text. Answers questions like: What info is in this text? Used for Natural Language Processing – NLP.
• Recommenders. Dimensionality reduction, collaborative filtering --think Netflix and Amazon. Predicts what someone might be interested in?
• Image Classification. Satellite Imagery. Classifies images with popular networks like ImageNet, MNIST, NASA Earthdata Search, and NOAA’s Image datasets.
• Anomaly Detection. Used for fraud detection, financial crimes, and credit card thefts. Identifies and predicts rare or unusual data points. Is this weird or unusual?

“Proposal For The Dartmouth Summer Research Project on Artificial Intelligence”, J. McCarthy, Dartmouth College. M.L. Minsky, Harvard University, N. Rochester, I.B.M. Corporation & C.E. Shannon, Bell Telephone Laboratories. August 31, 1955.

The individual created with coining the term “Artificial Intelligence -- AI” (rightly so) is the late John McCarthy. In 1955, John McCarthy, who earned a PhD in Math at Princeton, was an Assistant Professor in the Department of Mathematics at Dartmouth College. What is frequently misstated is the YEAR in which John McCarthy coined the term Artificial Intelligence.

Most (but not all) Google Search results erroneously state McCarthy coined the term in 1956. He didn’t! It bugs frequent AI/ML presenter, Rick Roche, that so-called AI experts and “reliable sources” misstate the year when McCarthy formally introduced the term AI. It is indicative of sloppy research habits. Dig deeper than superficial Google search results and alleged experts. Trust but verify! The year 1956 is actually associated with the “Dartmouth Summer Research Conference Project on AI”, held in August-1956 in Hanover, NH. Read it for yourself…

“Big Data and AI Strategies Machine Learning and Alternative Data Approach to Investing”, Quantitative and Derivatives Strategy Group, JP Morgan (May-2017)

This reviewer, Rick Roche, CAIA, is a  frequent presenter/commentator on “The Evolution of Machine Learning in Investment Management”. From 4Q 2017 through 4 Q 202, Rick has been the featured speaker on ML-Invt. Mgmt at 53 CFA Societies, Financial Planning Association (CFP) or CAIA Chapters and the New York Alternative Investment RoundTable (NY-AIR).

Rick considers JPMorgan’s “Big Data and AI Strategies Machine Learning & Alt-Data” a veritable “Bible” on machine learning applications used in investment management. As a communicator, he has often used the JPM “Classification of Machine Learning Techniques”  (refer to pages 16-19 and Figure 7 in particular). Don’t be scared-off by this 280-page report! Rick recommends you read the meat of this publication in page 1-134. Section IV is a “Handbook of Alternative Data” (pages 135 – 213). This Alt-Data Handbook is a compendium of Alt-Data Providers (as of 1Q 2017—many more Alt-data vendors as of 4Q 2020). Page 214 on is an Appendix which includes techniques for web crawling, ML packages and code & glossary.

Obviously there have been enormous advances in the use and application of ML algos in finance and alternative investment data since May-2017. JPM’s Quantitative and Derivatives Strategy Group themselves have published a second installment in this series. It’s a 322-page monster titled, “Big Data and AI Strategies 2019 Alternative Data Handbook”. Having digested the contents of both publications, Rick’s recommendation is to read the May-2017 “ML-Invt Mgmt Bible” first. Then if time permits, turn the pages on the JPM 2019 Alt-Data Handbook.

Rick's Originals: 

Rick Roche, CAIA, is a 39-year industry veteran, Chartered Alternative Investment Analyst (CAIA) and managing director at Little Harbor Advisors, LLC. Little Harbor Advisors is a sponsor of innovative alternative investments and volatility-informed trading strategies. Rick is a frequent speaker at CFA Societies and Financial Planning Association (FPA-CFP) Chapter events.

Rick's Bio