Evaluating AI-Based Hedge Fund Strategies

On 28 Mar 2017, I gave a presentation at the Battle of the Quants Conference in Frankfurt, Germany on how investors should assess AI-based investment strategies. The contents of the talk formed the basis for the white paper, "Artificial Intelligence Simplified: A Framework for Assessing AI-based Strategies."

This white paper presents a framework for assessing investing/trading strategies that are driven by Artificial Intelligence (AI).

Considerable hype surrounds the technology and that is to be expected given not only its capability today but what might be expected of AI in the future. This environment poses difficulties for investors and allocators who are trying to assess AI-based strategies in the present. Distinguishing hype from reality is the starting point. Then, investors must focus not on the technical aspects of the technology in all its various forms, but on what particular mispricing or inefficiency the strategy exploits, and the way in which AI enables the discovery and capture of such alpha.

Read the white paper here.


AI for Humanity

Have you met Robear? Robear is a Japanese robot being developed by Riken-SRK to care for the elderly. Currently 20% of Japan’s population is 65 or older – and the country faces the very real prospect of not having enough young workers to care for the old.  According to the Pew Research Center, in America only 13% of the population is over age 65, but expect that to double by 2050.  With respect to demographics, society has been much more focused on overpopulation over the past 50 years, but what if the real problem is population decline?

The trend is toward diminishing birth rates in Asia, Europe and the United States.  Birth rates are declining below the 2.1 live births per woman replacement rate.  In Germany, they’re around 1.36, Spain 1.48 and Italy 1.4. Russia and China are looking at their populations potentially declining by half. What about America?  The overall birth rate is a mere 646 per 1000 women. You don’t have to be a demographer to realize that’s simply not enough to sustain our population.

What does this mean for humanity? It is commonplace to talk about AI putting people out of jobs, with plenty of forums, conferences, and news articles on the subject.  But what about the specter of not having enough people to do those jobs in the first place? Among the myriad social implications of population decline, one of them is that we’re losing our workforce.  Who will stock the groceries, grow the food supply, haul the garbage, or deliver the mail?  Who will care for our increasing population of elders with no young people to replace them? And let’s not forget, we’re seeing large increases in the proportion of young people who will require care throughout their life because of various conditions.

Consider the contributions AI is making to healthcare. A team of Stanford researchers harnessed AI to detect skin cancer moles as accurately as a dermatologist. Melanomas represent fewer than 5% of all skin malignancies, yet they account for a majority of deaths related to this form of cancer. If detected early, the survival rate for melanoma is 99%. Imagine how many people could benefit from early detection. 

It is time to put AI into perspective. Forget about AI being the problem; AI can offer tools to enable people to work more efficiently and intelligently. We must learn to collaborate with AI, not fear it. Developments in AI can create opportunities and products for us to provide better care to those in need. At Greyfeather we're focused on what AI can do for investing, however, I believe AI will enable humanity to prosper in a number of areas given the challenges ahead. 

In Kevin Kelly's TED talk on the future of AI, he makes a compelling argument. He believes that AI can bring about a second Industrial Revolution. The future is not about AI undermining humanity, it's about humanity harnessing and collaborating with AI to do more and be more.

The AI Debate

Screen Shot 2016-11-03 at 9.49.15 PM.png

There’s a debate raging today about artificial intelligence -- between proponents and sceptics, mainly centering on how it will change the nature of the work that humans do in the economy.  There has been a lot of hyperbole and no shortage of “end-of-the-world-as-we-know-it” scenarios.  This debate has also made its way into investment management – but with a slightly different focus, with a key question being:

Can artificial intelligence power profitable trading/investing strategies?  

I don’t even need to tell you how I would answer that question, but believe it or not, many have come out strongly against trading/investing through the use of AI – and unfortunately, much of the media coverage has been sensationalist, without the technical substance to demonstrate what this technology can and cannot do.  The main thrust of the criticism is that AI is a commodity, deployable by any organization and will provide the same answers to anyone who uses it.  Ergo, if everyone uses it, then it cannot possibly provide a trading edge. 

The problem with this logic is that AI is not uniform – there are countless techniques, model configurations, design choices, and vastly different data.  What data researchers use and how they process it can make all the difference in the world.  How one trains a model, and what inputs are going to be used for training is going to be completely different across asset classes, applications and the interests of different managers.  Different market participants have different interests – e.g. compare market-makers to long-only mutual funds – they look at opportunities differently and are going to deploy technology in different ways.  Could it be profitable in both contexts?  Absolutely.  Like computers in general, AI is a tool, which depends critically on who is wielding it and how.  Not all machine-learning algorithms are created equal.

I think what most invigorates the opposition to AI is a belief that AI is threatening human analysis and contribution – knowledge work in general.  This is not the case yet, and it is not likely any time soon.  Building AI models and processing data, takes a lot of human toil – just ask the team at Greyfeather.  Most importantly, though, machines are not self-aware even if it might be possible to fool humans into thinking that they are.  But let’s be honest about human limitations.  Any given trader, PM, investment manager - how many opportunities can they intelligently look at? Can we reasonably expect them to respond to important signals systematically and effectively?  And do so optimally?

The debate on AI and investment management is really the “Efficient Markets” debate masquerading as something else.  The line of reasoning used by AI sceptics is quite similar to those who refuse to acknowledge the drawbacks of the efficient market hypothesis, i.e. that market prices reflect all relevant information about a security.  Empirical evidence, both published and unpublished, along with recorded market history – the irrational bubbles and severe crashes, makes this argument untenable from our perspective.  We certainly don’t mind though that much of academia along with other practitioners continue to believe in the EMH.  I say believe, because it is an act of faith at this point. 

Dissecting a Neural Network

Today, let's look inside a neural network. I'm Joanna Sandretto, the Lead Developer here at Greyfeather, and building on our last couple posts - let's go a step further. 

A good way to understand neural networks is to look at a simple example that implements the XOR operation. XOR is the logical operation for “exclusive or”, which determines whether one, and only one, of two inputs is true. That is, given inputs x and y, the result is true only if x or y is true but not if x and y are both true or both false. We can represent the XOR operation in a truth table as follows where 1 represents true and 0 represents false:

Blog Ideas-13.jpg

The figure below shows a simple neural network with an input layer, one hidden layer and an output layer. The first layer has our two inputs x and y along with a bias unit that has a value of 1. The hidden layer has two nodes plus a bias unit with a value of 1. The threshold is 0, so the neuron will fire when the neuron’s value is greater than 0. The bias units always fire because its value is 1, which is greater than 0. Each node often has a complex function, but for simplicity here each node sums all the input values and compares the result to a threshold. The numbers along the edges connecting the neurons are the activation values that are passed to the next layer when the neuron fires.

As a concrete example, say x is true and y is false. Here x will be 1 and y will be 0, so the neuron associated with x will fire and pass its activation values on to the next layer while y will not. In the second layer, both values are above our threshold 0, so both neurons will fire and pass an activation value to the output layer. The value of the neuron in the output layer is -20+15+15=10, which is greater than 0, so the output of the network is 1.

Now let’s say x is true and y is true. In this case, both x and y will fire and pass their activation values to the next layer. As expected, the result is 0.

The XOR operation is an excellent example of why neural networks are helpful in solving complex problems. As problems become more complex, solving the problem often requires combining multiple functions or operations. Neural networks allow us to do just this. In the case of XOR, we need to check for:

x OR y

x AND y

not x AND not y

If all of these operations can be combined in a network with a single hidden layer, imagine what is possible in a deep neural network with multiple hidden layers...

Types of Neural Networks

There are many types of neural networks, but we can characterize most neural networks as either feed-forward networks or recurrent networks.

In feed-forward neural networks, the information travels in a single direction. When a neuron fires in a feed-forward network, the activation value is always passed to the next layer and is never allowed to pass an activation value back to itself or to a previous layer. Sometimes the network will have only a few hidden layers, but other networks will have many hidden layers. Neural networks that have more than one hidden layer are considered Deep Neural Networks (DNN).

Hidden Layers in a DNN

Hidden Layers in a DNN

There are different types of feedforward neural networks that can be used for different problems or types of data, including Multilayer Perceptrons (MLP) and Convolutional Neural Networks (CNN). MLP are networks with multiple hidden layers with each node in a layer connected to every node in the next layer. The figure above shows the architecture of an MLP. CNN are networks where the inputs are assumed to be images. This assumption allows the network architecture to be modified to make processing images more efficient.

In contrast, recurrent neural networks have connections that pass data towards the output layer as well as connections that pass data back to previous layers. These connections create loops which act as a sort of memory of previous inputs. The resulting “memory” feature of recurrent neural networks make them useful in processing sequences or time series because information about previous inputs can be used to process later inputs. A recurrent neural network may look like:

Recurrent Neural Network

Recurrent Neural Network

Because predictions in finance often involve time series data, a recurrent neural network can be an ideal neural network architecture to use to make financial predictions. 

How Neural Networks Work - A Primer

I'm Joanna Sandretto and I work as the Lead Developer at Greyfeather Capital. I also happen to be Matt's sister, in case you are wondering. I am a computer scientist whose focus is artificial intelligence. I think it is important for people to understand what neural networks are, and what they can (and cannot) do - so let's start at the beginning.

Artificial intelligence (AI) may seem like the new cool “thing” in computer science with companies using AI to build assistants like Siri or Cortana, predict what shoppers will buy, and even drive cars, but computer scientists have been researching artificial intelligence for decades. Machine Learning (ML) and Deep Learning (DL) are areas of AI that focus on systems that can learn and improve without human intervention. Neural networks are one way to build systems that learn.

Artificial neural networks are an AI technique that computer scientists have been researching since the 1940s. In fact, Marvin Minsky and Dean Edmonds used 3000 vacuum tubes and a surplus automatic pilot mechanism from a B-24 bomber to build the first neural network computer in 1950 (See Artificial Intelligence: A Modern Approach by Russell and Norvig).

Artificial neural networks are based on the structure of a human brain. In basic terms, each neural network consists of neurons, or units, that are connected to each other. Each unit is associated with some function that will map multiple inputs to an output. For example, a network with two inputs and a single layer might look like the following:

A neural network with two inputs and a single layer

A neural network with two inputs and a single layer

The function that inputs are passed through is referred to as the activation function.  The activation function controls what value gets passed to the next layer.  A neuron is said to “fire” when it passes a value, referred to as an activation value, to the next layer.  In some networks, neurons will fire if the input value is above some value and otherwise pass no value to the next layer. In more complex networks, neurons can fire with different strengths and the activation function will calculate the value that should be passed to the next layer. 

There are many types of networks, but in most cases, the units are organized into layers with each layer consisting of several neurons. Networks have different number of layers and different number of neurons in each layer. The basic goal of the network is to create a mapping between the inputs and the desired outputs, which will allow the network to accurately predict the output given certain inputs. 

The first layer is an input layer. The number of neurons in the input layer is defined by the number of features or inputs chosen for the data. As a crude example, let’s say that we wanted to investigate to what extent leverage levels are associated with overall stock performance.  We could have the output be the 3-month cumulative return and the input layer could be the past 12 quarters of debt/equity ratios.  In this case, the input layer would have 12 neurons, one for each quarter’s debt/equity ratio. 

After the input layer, the network will have one or more hidden layers. The hidden layers will transform the data using various functions. Different networks and different types of data will require different types of functions. These layers are creating the mapping from inputs to outputs. 

As data passes through the network, the weights associated with the function will be adjusted based on certain rules, this is the learning part of the neural network.  A popular way to update weights is using backpropagation.  In a network using backpropagation, the error of the network is calculated and fed into an optimization function, which then adjusts the weights of the neurons accordingly.

The End of Code: A Revolution in Computer Science, Part III

Technological advancement is changing our world at breakneck speed, necessitating businesses in all industries all over the world to keep pace – or be left behind the competition. Machine learning via deep neural networks have created a seminal moment in the field of artificial intelligence (AI).

Andy Rubin, founder of Playground Global, a company focused on machine learning investment, notes, “People don’t linearly write the programs.  After a neural network learns how to do speech recognition, a programmer can’t go in and look at it and see how that happened.  It’s just like your brain.  You can’t cut your head off and see what you’re thinking.”  Greyfeather’s neural networks work the same way, but instead of Rubin’s example of learning to recognize speech, they learn to recognize features of stock data that are indicative of future performance.

Wired Magazine’s Jason Tanz notes the importance of this paradigm shift, “If the rise of human-written software led to the cult of the engineer, and to the notion that human experience can ultimately be reduced to a series of comprehensible instructions, machine learning kicks the pendulum in the opposite direction.  The code that runs the universe may defy human analysis.”

With this pendulum swing comes the opportunity for technically advanced entrepreneurs, and those who invest in them, to achieve lucrative returns.  Sebastian Thrun, a former Stanford AI professor and creator of Google’s self-driving car, says, “Neural nets had no symbols or rules, just numbers.  That alienated a lot of people.”  It does not alienate Greyfeather Capital.  We have seized on this opportunity to be among the first to market with this cutting edge technology in order to provide our investors with a new edge.

Steven Hawking echoed the sentiments of Elon Musk, Bill Gates, and others when he wrote, “One can imagine such technology outsmarting financial markets, out-inventing human researchers, out-manipulating human leaders...”  While the potential applications of this technology are nearly limitless, Greyfeather Capital has completely focused on developing AI to select stocks with the highest probability of out-performance. The future of investment management, as Steven Hawking describes it, is now.

The End of Code: A Revolution in Computer Science, Part II

When analyzing investment management, or any industry for that matter, an important question to ask is – Why is now the right time to invest in a technological paradigm shift?

To answer this, investors should heed recent technological breakthroughs and observe industries outside traditional investment management.  Wired Magazine’s Jason Tanz notes, “Over the past several years, the biggest tech companies in Silicon Valley have aggressively pursued an approach to computing called machine learning…. This approach is not new – it’s been around for decades – but it has recently become immensely more powerful, thanks in part to the rise of deep neural networks.”

Deep neural networks, which power Greyfeather’s stock selections, have spawned a renaissance in the field of artificial intelligence.  These networks have proven the missing link allowing computer software to evolve beyond explicit, step-by-step instructions to a state in which it learns in a manner similar to the human brain.

Industry has certainly taken notice!  Google has used this technology to make the self-driving car possible.  Netflix uses this technology to predict with movies and shows you are likely to enjoy.  Computers are now able to consistently defeat human beings in complex games, such as the ancient game of Go. Andy Rubin, a computer programmer and entrepreneur, has started a new company, Playground Global, which invests solely in machine learning startups. Even the National Football League has invested heavily in artificial intelligence – not wanting to be left behind the technological power curve!

In the investment management industry, AI is not being deployed to an extent commensurate with its capability.  Greyfeather Capital has been quick to recognize the technological paradigm shift, has the in-house talent to capitalize on it, and is applying it to the industry. Just as Jeff Bezos, seeing the exponential rise in internet usage, left a lucrative quant hedge fund job to found Amazon.com, we at Greyfeather Capital are seeking forward-thinking investors to pioneer a new, rigorous, and technologically driven era of investment management.

The End of Code: A Revolution in Computer Science, Part I

I'm Jeff Payne and I work as the Chief Operating Officer at Greyfeather Capital.  One of my struggles in this role has been finding the right words to convey the unique and revolutionary nature of the technology that powers Greyfeather’s strategy. 

That being the case, you can imagine my excitement when I found Wired Magazine’s article “The End of Code” written by Jason Tanz of the June 2016 issue. This article explains, in layman’s terms, the type of artificial intelligence that Greyfeather is using and how it is a radical departure from the computer code used by many of today's quantitative hedge funds. Over the course of three posts, I will highlight some key points from that article that provide context to the manner in which Greyfeather will disrupt the field of investment management. 

For decades, computer science has been ruled by the programmers.  Those who write the best code reign supreme.  The process is simple to articulate, but takes extreme talent to execute – the coder writes rules in a computer programming language and the composite logic derived from those rules dictates outcomes.  The collection of these rules is often called a “black box” from which outputs seem to magically appear.  Furthermore, it was long assumed that the more elegantly written code you write, the more exactly you could approximate the functionality of the human brain.

However, recent developments in the understanding of the human brain, have caused computer scientists to rethink that presumption.

Wired Magazine’s Jason Tanz notes that “the brain [isn’t] a black box at all.”  It turns out, in the mid-1950’s, psychologists began to argue that, “people… were not just collections of conditioned responses.  They absorbed information, processed it, and then acted upon it.  They had systems for writing, storing, and recalling memories.”

It is this revelation that has allowed scientists to create the modern field of artificial intelligence (AI), bring the discipline of machine-learning to its forefront, and make deep learning neural networks its most potent weapon.

The contrast in this computer science paradigm shift cannot be understated.  As Tanz notes, “If in the old view programmers were like gods, authorizing the laws that govern computer systems - now they’re like parents or dog trainers.”

“If you want to teach a neural network to recognize a cat… you don’t tell it to look for whiskers, ears, fur, and eyes.  You simply show it thousands and thousands of photos of cats, and eventually it works things out.”

This is Greyfeather’s approach.  If you want to teach a neural network to recognize a stock likely to outperform the upcoming month’s median return, you don’t tell it to look for momentum characteristics, attractive valuations, etc.  You simply show it copious market and fundamental data and the neural network learns to make predictions that no team of stock analysts can match.

The application of this revolutionary technology is why Greyfeather Capital is not just another quant hedge fund.  Rather we are a pioneer of artificial intelligence in the field of investment management.  


Thinking Outside the Black Box

One of the biggest criticisms, and rightly so, of quantitative trading is that any given black box strategy, much like a set mathematical function, is rigid and mechanical.  It isn’t a perfect fit and it can’t account for all scenarios and market movements.  Artificial Intelligence, involving deep learning and neural networks, can address this shortcoming.  Where traditional data mining fails, AI succeeds.

You may have heard (or read) about big data, machine learning and neural networks at some point.  Let’s keep it simple - big data is just that, a collection of data. How do you capture, curate, search, store, update and analyze all of it?  We look at equity market data and financial performance metrics for over seventy-five years for all companies listed on North American exchanges.  That’s a lot of information, with more being recorded every day. So how do we process that information?

We could have hired an army of analysts to sift through the data, but that would be disorganized and inefficient at best.  Instead, we apply a smarter, more efficient and more effective method. We use our artificial neural networks, which are algorithms inspired by the human brain, designed to learn the complex, non-linear relationships between sets of inputs and the desired output (prediction).  It is a form of machine learning, where a machine (or computers) process the information, analyze it and develop an opinion on what they see. No human team, however large or talented, can compete with an AI-based process.

Here at Greyfeather, we don’t like to think of ourselves as quant traders.  When we talk to others in the industry, the term doesn’t really explain what we do. Sure, our strategy is quantitative, but our technology and research approach is what makes us different. Traditional quant traders tend to computerize trading execution of a strategy that is typically either trend-following or mean-reverting and based on technical factors (price/volume) or fundamentals.  We are an AI-based trading strategy in which we seek to have all relevant factors assimilated and optimize predictions.

Quant trading was an evolutionary advancement from discretionary trading that brought computerized discipline to strategy execution. AI represents a further advancement from quant trading that brings supercomputer learning to understanding the markets - an advantage no quant or discretionary trader enjoys.  This technology renders the "black box" obsolete.  We are replacing the outdated "black box”, by bringing the latest developments in AI to equity long/short investing strategies on Wall Street. 

Why Artificial Intelligence is the Future of Wall Street

Screen Shot 2016-08-25 at 8.59.04 PM.png

Today, it is easy to assume that all frontiers have been explored on this earth but is that really true?  While much of the natural world has been surveyed and catalogued, in the realm of computing and algorithm development, in particular, artificial intelligence, we are just beginning to scratch the surface.  My name is Matthew Sandretto and I have been working at developing trading algorithms for futures, equities and options for over a decade.  I’ve had some great successes and of course, setbacks too, but for the past couple years, I’ve been completely focused on developing smart trading algorithms based on artificial intelligence.  What if you could teach a computer to read the ticker tape for thousands of stocks simultaneously, to learn from that data and then make trades based on the learning that has taken place?

Really this isn’t such a radical idea.  On a smaller scale, it is the very same thing that Jesse Livermore, one of the most notable traders of the 20th c., claimed that he did within his own mind.  While the brain of Jesse Livermore was undoubtedly successful at learning – e.g. he saw an eerie similarity in the market conditions of 1929 to those of 1907, which enabled him to be short prior to the crash; an AI-based algorithm offers many advantages:

  • Analyze thousands of tradeable instruments
  • Long memory
  • Able to assign probabilities rigorously
  • Not subject to emotions and the caprice of human behavior (Our psychological needs can influence decisions, in many cases causing sub-optimal outcomes)
  • No longer data/processor limited with virtual machines/cloud computing
  • Artificial intelligence holds the promise of being able to process all of this information, learn from the data and evolve to be the most effective trader on the Street

The reality is that the market is and probably always will be influenced by human participants.  Should we expect human participants to be able to adequately analyze this environment that is influenced both by their trading decisions and by their act of analyzing it? (see Soros’s principle of Reflexivity) In Charles Mackay’s book “Extraordinary Popular Delusions and Madness of Crowds” he writes: “Men, it has been well said, think in herds; it will be seen that they go mad in herds, while they only recover their senses slowly, and one by one.” 

Artificial intelligence is the future of trading because it is able to observe the market, and recommend objective decisions learned from its behavior.