Financial Machine Learning - A Practitioner’s Perspective by Dr. Ernest Chan

Introduction

In this section, the speaker introduces himself and his background in finance and machine learning.

Speaker's Background

• The speaker is Ben, one of the education directors on Quant.

• The guest speaker is Ernie Chan, a managing member of QTS Capital Management and founder of PredictNow AI.

• Ernie has authored several books on quantitative trading and made numerous contributions to the quantum finance community.

Financial Machine Learning

In this section, the speaker talks about his favorite topic - financial machine learning. He discusses his history in machine learning and why it has been difficult to extract value from applying it to finance.

History in Machine Learning

• The speaker has a long history in the field of machine learning starting at IBM TJ Watson Research Center.

• He found that until recently, it had been extremely difficult to extract value from applying machine learning to finance.

Value in Applying Machine Learning

• It was only a year and a half ago that he found value in applying machine learning to finance.

• He continues to find value by applying machine learning in a way that most people are not doing.

Simple Models vs. Complex Models

• The speaker has always advocated for simple models as the most likely path to profit.

• Simple models mean single factor or linear models which have worked well since the 1980s.

• However, with more people entering quantitative trading, these simple models become less profitable due to competition.

Overfitting Issue

• Another reason why machine learning was not particularly successful in finance was due to overfitting issues.

• In simple linear models, there are just a couple of parameters that you have to adjust which can lead to overfitting.

Introduction to Machine Learning in Finance

In this section, the speaker introduces machine learning and its application in finance.

Machine Learning Models

• Financial time series data has limited data points to support fitting all parameters of a machine learning model, making it easy to overfit.

• Advances in avoiding overfitting have been made through techniques such as dropout and random forest.

• Feature selection is a tool used in interpretable machine learning models that identifies important inputs to the model.

Transparency in Machine Learning

• Black box trading can be uncomfortable for money managers who want to understand why trades are being made.

• Interpretable machine learning models aim to identify important features that lead to conclusions, even if the exact reasoning behind those conclusions is not fully understood.

Conclusion

• The use of machine learning models in finance requires transparency and interpretability for effective decision-making.

Feature Selection

In this section, the speaker explains how feature selection is used to rank and identify important variables in machine learning models. He also discusses how this technique can be used for risk management and capital allocation.

Importance of Feature Selection

• Feature selection helps to identify the most important variables in a machine learning model.

• It provides an intuitive and interpretable understanding of the model's behavior.

• This information can be used to explain model performance to investors.

Use Cases for Machine Learning in Finance

• Machine learning can be used for risk management and capital allocation.

• It is not effective as a primary signal generator because many other programs are competing for the same predictions.

• The financial market is constantly evolving, making it difficult to predict with accuracy.

Risk Management and Capital Allocation

In this section, the speaker explains why he believes that machine learning should primarily be used for risk management and capital allocation rather than as a primary signal generator.

Limitations of Machine Learning as a Primary Signal Generator

• Many other programs are competing for the same predictions, making it difficult to achieve accurate results.

• The financial market is constantly evolving, which makes it challenging to predict with accuracy.

Advantages of Using Machine Learning for Risk Management and Capital Allocation

• Private information can be used to gain an edge over competitors.

• Machine learning can help identify patterns that may indicate increased risk or opportunities for investment.

Applying Machine Learning in Finance

In this section, the speaker discusses some of the challenges associated with applying machine learning techniques in finance.

Challenges Associated with Applying Machine Learning in Finance

• The financial market is constantly evolving, making it difficult to predict with accuracy.

• Machine learning programs are competing for the same predictions, making it challenging to achieve accurate results.

• The financial market exhibits reflexivity, which means that predictions can impact market behavior.

Advantages of Using Machine Learning in Finance

• Machine learning can be used for risk management and capital allocation.

• Private information can be used to gain an edge over competitors.

Machine Learning vs Traditional Quantitative Trading Strategies

In this section, the speaker discusses the differences between traditional quantitative trading strategies and machine learning-based strategies.

Parameters and Predictors

• Traditional quant strategies have few parameters and predictors.

• Machine learning-based strategies have numerous predictors due to their ability to handle non-linear dependencies of predictors.

Data Modeling

• Traditional quant models use prices and fundamentals as data.

• Machine learning models can model alternative data that is difficult to build a mathematical model on.

Non-Linearity

• The importance of non-linearity in machine learning models is that not all features will be used, making it easier to get rid of some features.

• Non-linear dependencies of predictors can be easily handled only with non-linear models like random forests.

Advantages and Disadvantages

• Traditional quant models are easy to understand but easy to replicate, leading to decay.

• Machine learning models are unintuitive, opaque, and may be black boxes but less likely to decay because they are difficult for two people to build the same model.

Alpha Decay and Traditional Quantitative Models

In this section, the speaker discusses alpha decay and traditional quantitative models.

Alpha Decay

• Alpha decay is less likely in traditional quantitative models.

Traditional Quantitative Models

• Traditional quantitative models have a big problem: they only tell you to buy or sell without indicating the likelihood of success.

• It is hard to assess the statistical significance of backtests in traditional quant models because it is hard to simulate the market realistically.

Machine Learning Models for Trading

In this section, the speaker discusses machine learning models for trading.

Likelihood of Success

• Machine learning models can provide a probability of success for trades, allowing traders to allocate capital accordingly.

Randomness and Backtesting

• Unlike traditional quant strategies, machine learning-based strategies are not deterministic due to intrinsic randomness. This allows for multiple backtests with different results, which helps assess statistical significance.

• It is difficult to create an error bar for sharp ratio in traditional quant strategies due to limited predictors and linearity. However, machine learning models allow generating multiple backtests by switching random seeds.

Applying Machine Learning to Trading

In this section, the speaker outlines three steps involved in applying machine learning to trading.

Three Steps

• The first step is financial data science.

• The second step involves solving the machine learning problem.

• The third step involves creating a trading strategy based on the output from step two.

Financial Data Science Challenges

In this section, the speaker discusses the challenges of financial data science and emphasizes the importance of human intelligence in solving these challenges.

Problems with Financial Data

• Financial data is prone to numerous problems, even from reputable vendors.

• Complete Stat Data, which provides company fundamentals, distributes a version containing look-ahead bias to universities for free.

• Restated financial data can cause issues when used in machine learning models.

• Sentiment data can be unreliable due to companies re-engineering their processes to make themselves look good.

Engineering Features

• After obtaining good data, it must be engineered into useful features, which requires domain expertise.

• Machine learning cannot solve the problem of financial data science; it requires human intelligence.

Trading Strategy Construction

• The third step is constructing a trading strategy based on predictions generated by machine learning models.

Putting Predictions into a Coherent Strategy

In this section, the speaker discusses the steps required to put predictions into a coherent strategy and assess its statistical significance.

Creating a Coherent Strategy

• Having a prediction is not enough; it needs to be turned into a coherent strategy.

• Backtesting the strategy is necessary to assess its statistical significance.

• There are many platforms available for backtesting, such as QuantConnect and Quantopian.

Human Intelligence in Strategy Creation

• While there are many technological solutions for backtesting, human intelligence is still required in creating strategies.

• Converting predictions into portfolios follows standard routines outlined in finance textbooks.

Difficulties of Financial Data Science

This section covers the difficulties of financial data science and how they differ from traditional data science.

Stationarity of Features

• Most financial data is non-stationary and cannot be used as features without being converted first.

• The process of converting non-stationary features into stationary ones can be complex and requires domain expertise.

Machine Learning Models for Finance

• Linear and logistic regression models are too simple for most financial applications.

• Deep learning models like deep neural networks have too many parameters to fit with limited financial data.

• Random forests are commonly used because they capture non-linearity well without being overly complex.

Meta Labeling

In this section, the speaker discusses meta labeling as an effective technique for using machine learning in finance.

Predicting Profitability Instead of Market Movements

• Rather than predicting market movements, predict whether your own strategy will be profitable.

• This unique dataset can only be predicted by you and provides more accurate results than trying to predict the market.

• This technique is called meta labeling and is covered in more detail in Marcos Lopez de Prado's textbook.

Meta Labeling Applied to Finance

In this section, the speaker talks about the importance of meta labeling in finance and how it was applied to their trading strategy called Tail Reaper. They also discuss the role of machine learning in detecting risks and stopping trading strategies.

Importance of Meta Labeling

• Meta labeling is important in finance.

• It was applied to their trading strategy called Tail Reaper.

Role of Machine Learning

• Machine learning detected risks on February 1st when nobody thought that the virus would have any impact on the world economy.

• The machine learning model advised them to stop hedging terrorists a few days before Pfizer's announcement about a vaccine.

• The first announcement of the vaccine caused a huge move in the financial market that wiped out many momentum funds.

• Their machine learning model advised them to stop trading that strategy, which saved them from suffering tremendous losses.

Feature Selection and Predicting Macro Variables with Machine Learning

In this section, the speaker discusses feature selection and its importance in explaining losses to investors. They also talk about using machine learning to predict macro variables accurately.

Feature Selection

• Feature selection is important if you want to explain losses to investors.

• A pre-print paper has been accepted for publication at the Journal of Financial Data Science about feature selection.

Predicting Macro Variables with Machine Learning

• A white paper has been published on how to use machine learning to predict macro variables accurately.

• An alternative data company provided data for this study, which helped predict non-farm payroll surprises with surprising accuracy.

Recommended Reading Materials

In this section, the speaker recommends reading materials for those interested in learning more about machine learning and artificial intelligence.

Recommended Books

• The speaker recommends reading his book on machine learning or the textbook by Murphy called "Machine Learning: A Probabilistic Perspective."

• Before reading Dr. Michael Pracitce's book, the speaker suggests starting with his book or Murphy's textbook.

Other Resources

• The speaker recommends checking out their white papers, blog posts, and presentations on BJEc for more information on machine learning in finance.

Introduction to Fundamental Data in Machine Learning

The speaker introduces the concept of incorporating fundamental data into machine learning models and explains how machine learning has an advantage over classical statistics when it comes to handling categorical data.

Incorporating Numerical and Categorical Data

• Numerical data, such as earnings per share or PE ratio, can be easily incorporated into machine learning models.

• However, some fundamental data may be categorical in nature, such as whether a company declared a dividend or stated its earnings last quarter. Linear regression models cannot handle categorical data.

• Machine learning models have the ability to condition on any variable, whether it is categorical or continuous. This makes it easy to mix and match real value and categorical data in the model.

Limitations of Machine Learning in Finance

The speaker discusses the limitations of using machine learning in finance and why it cannot replace human traders entirely.

Why Machine Learning Cannot Replace Human Traders

• 80-90% of financial machine learning process is in financial data science which requires deep domain expertise. Creating features is an intensely human process that cannot be replaced by machine learning alone.

• Blind worship of deep learning is detrimental to financial machine learning. Deep learning cannot solve everything if there isn't enough relevant data available.

• Directly applying deep learning to predict returns is unlikely to succeed. Most successful startups find a niche rather than directly competing with well-funded organizations.

Conclusion

The speaker concludes by answering a question about where he sees machine learning going in finance.

Future of Machine Learning in Finance

• The speaker highly doubts that machine learning will ever be good enough to replace human traders entirely.

• Machine learning can be a useful tool in finance, but it cannot replace the expertise and intuition of human traders.

Capital Allocation and Expected Return

In this section, the speaker discusses capital allocation and expected return. They explain that capital allocation involves using popularity as an input to determine how much capital to allocate. The expected return is a key input in traditional models, but historical returns are not a good judge of future returns. Machine learning can provide a more sophisticated expected return as an input to the model.

• Capital allocation involves using popularity as an input to determine how much capital to allocate.

• Traditional models use historical returns as the expected return, which is not a good judge of future returns.

• Machine learning can provide a more sophisticated expected return as an input to the model, resulting in proper portfolio allocation.

Static vs Dynamic Probabilities

In this section, the speaker discusses static probabilities versus dynamic probabilities in classical statistics and machine learning.

• Classical statistics provides static probabilities of distribution every day, while machine learning provides different probabilities every day based on varying inputs.

• In classical statistics, error bars do not depend on independent variables; however, in machine learning, error bars vary with inputs. This is desirable for machine learning but problematic for classical statistics.

Deep Learning and Feature Selection

In this section, the speaker discusses deep learning and feature selection in financial models.

• Deep learning may be useful for price series analysis but does not eliminate the need for creating different types of time series features or cross-sectional diversity of inputs necessary for making predictions successfully.

• Feature selection requires domain expertise, and new deep learning methods will not help with that.

Reinforcement Learning in Financial Models

In this section, the speaker discusses reinforcement learning in financial models.

• Reinforcement learning has a place in financial models but does not eliminate the need for domain expertise or feature selection.

• The time scale of reinforcement learning depends on the problem being solved.

Reinforcement Learning and Trading

In this section, Dr. Chan discusses the application of reinforcement learning in trading.

Reinforcement Learning in Trading

• Dr. Chan explains that easy trading can be achieved through reinforcement learning.

• Reinforcement learning can react to people placing orders on the order book, making it useful for shorter time scales.

• There is no convincing evidence that reinforcement learning works in longer time scales.

Predict Now AI

• Dr. Chan's company, Predict Now AI, provides a tool for no-code financial machine learning.

• The tool is a great resource for those who are not interested in the programming side of things and want the expertise of someone like Dr. Chan and his company.

Overall, Dr. Chan's insights suggest that reinforcement learning can be useful for short-term trading but may not work as well for longer time scales. His company, Predict Now AI, offers a no-code financial machine learning tool for those who want to benefit from his expertise without having to do any programming themselves.