AI and deep learning are set to revolutionize the world, but can they tame the wilds of market behavior?
Sarah could sense something was amiss. It was a little too quiet. The hair on the back of her neck felt as if someone, no, some thing was watching her and it left her unnerved.
Trekking the better part of the last four days across the dystopian landscape that was once the bustling metropolis of Los Angeles, Sarah knew that she had to find water and fast.
But she wasn’t the only one who knew that she needed water — “they” were watching as well — “they” knew she needed water.
Looking down at her son John, whose lips were cracked and bleeding from dehydration, she knew that if she didn’t get to fresh water soon, “they” wouldn’t need to finish her off, dehydration would.
As Sarah made the 100-yard dash to the water pumping station, the glint from sunlight reflecting against highly polished steel caught the corner of her eye.
Running for cover, a hail of bullets rained down on her position as a trio of T-800 killbots ambled towards her position, scanning and firing as they made their way across the scarred landscape.
And while the T-800s could run indefinitely without water, Sarah and her son John could not — something which Skynet — an artificial neural network-based conscious group mind and artificial general superintelligence system, or AI for short — was acutely aware of.
Ask most people about the dangers of AI and deep learning and you’ll find that their immediate and most accessible reference will be to the Terminator series of movies.
And while it is entirely possible that a future AI will one day develop semi-sentience and take over the world with killbots, the more immediate danger from AI is not its current capabilities, but the overselling of those capabilities.
Hardly a day goes by without investors being told that AI will revolutionize investment management.
After all, if AI can drive cars (badly), beat world masters at chess and auto-fill our search queries, surely it can do something as pedestrian as make money consistently.
As attractive as this investment spiel looks, it doesn’t paint the entire portrait when it comes to AI and investing.
Money for nothin’, and chicks for free
— “Money for Nothing” by Dire Straits off the album “Brothers in Arms”
And while many may think of AI as some form of robot life form or even C3PO from Star Wars, it really isn’t anything more than statistics.
Because almost every sentence that uses the term “artificial intelligence” or “AI” can be substituted quite readily with the term “statistics” — it just sounds a lot less sexy.
Take for instance this statement,
“China intends to lead the world in AI.”
A statement like that conjures up Beijing building up robot armies that will one day dominate the world. But what if the word “statistics” was inserted in the same statement instead?
“China intends to lead the world in statistics.”
Meh. Doesn’t seem quite as scary does it?
Yet in reality, that’s what AI is — statistics.
The caricature is that an investor should be able to take a bunch of data (preferably exotic “alternative” data), feed it into some sort of “neural net” (maybe an early version of Skynet) and this “neural net” will become a literal money-printing machine, disgorging wealth for its creators beyond the dreams of avarice.
Such bare-faced claims understandably frustrate statisticians who know that AI today is simply predictive probability.
And while AI may be great at things like chess, it’s terrible at things which are dynamic or which have no clearly defined outcomes.
To begin with, machine learning requires a clear goal and within a controlled environment, AI has performed exceedingly well, beating humans at everything from chess to Jeopardy.
But the goals of investment and finance are less well-defined.
Just ask your average investor why they invest their money and you’ll get as many answers as you have investors.
“To get a better return than bank deposits.”
“To retire early.”
“To build a rainy day fund.”
“To pay for my kids’ college tuition.”
And these investment goals aren’t static either as they change according to an investor’s circumstance and age.
But what if the investment goal was simply to “make more money”?
Seems simple enough — so would a ten times leveraged position on the S&P500 that will make ten times the returns over time — but at substantially more risk be preferable to say buying a T-Bill? How much risk can an investor accept? And over what timeframe? And how much is more?
Are risk-adjusted returns the goal or is diversification also a necessary criteria?
Simply telling AI to “make more money” doesn’t specify the level of risk that is acceptable or the methodology to take.
But what AI can do, is sift through the mountains of data to suss out specific opportunities which human investors have identified according to preset parameters.
Unfortunately, because of the internet, the democratization of data means that investment managers are feeding the same data into AI — opportunities that are obvious and exist are very quickly discovered and traded out of existence.
It’s relatively straightforward to instruct AI to find specific opportunities that the data reveals, what is far more difficult is gearing programs to find what hasn’t yet been discovered.
“Don’t tell me what everyone knows. Tell me what no one knows,” is hardly a programmable instruction.
And even if it were possible to specify the desired returns, it leads to a secondary problem, if the opportunity is small or shortlived, or hard to exploit or scale meaningfully for a billion-dollar fund.
AI is not a crystal ball, predictions are probabilistic at best and misleading at worst.
Because AI systems have to be “taught” from past data, to determine patterns and probabilistically make an approximation of the future, it is no different from any other systematic or discretionary investment process — think of it as a “data-driven hunch” or “AI intuition.”
Unlike chess and Go, in finance, the past is often not a good indicator of the future. To use a statistical term, data in finance is “non-stationary.”
In the majority of fields where AI excels, the data tends to be stationary and within defined statistically repetitive boundaries.
Whether it’s movie recommendations or scanning X-rays to spot abnormalities in patients, data gleaned in those circumstances tends not to change all that much even over time.
For instance, Google’s legendary AlphaGo, knew exactly where the pieces on the board were and that doesn’t change with each new game played.
And nobody unless they were schizophrenic, chooses 19 random movies for every one movie they like and then expect that Netflix would come up with a good recommendation.
But consider that until 2020, nobody had even heard of the coronavirus, let alone factored it into its impact on markets. There are simply too many black swans in financial markets for AI to cater for.
And while it is possible to train an AI system to respond rapidly to events, that also means that it has to “learn” a new model based on a very short history, reducing the amount of data that the system can learn from and which increases the possibility of extreme overreactions, out of proportion to the magnitude of the stimulus.
Consider the recent thousand-point fall in the Dow — an AI programmed to react quickly may have then gone all-in to short the market and would have been caught off guard during the subsequent quick reversal and recovery.
Situations like that have already happened in the past, with numerous flash crashes and equally instantaneous recovery of the past decade. Robots can and will often overreact, sometimes to nothing more than their own shadows.
Then there’s the issue of the quality of the data fed to AI — it’s messy at best. Although what moves markets is not entirely random — the signal to noise ratio of the data received is exceedingly low.
There’s a reason why former Federal Reserve chairman Alan Greenspan coined the term “irrational exuberance” to describe market behavior.
And with the rise of so-called “alternative data” more smoke and mirrors has been added to the already confusing AI light show.
It is possible to use sophisticated techniques to reduce the effect of randomness in finance, but the quality of the data makes it challenging to apply to machine learning.
But that doesn’t mean that we won’t ever be able to build our very own Skynet.
While it is unlikely that financial AI tools will create new scalable sources of returns, it is proving extremely useful in far more mundane tasks, freeing up us humans for what we’re good at — intuition and creating neural linkages, spotting patterns otherwise indiscernible and scoping out opportunities that are statistical blindspots.
AI is very good at cleaning data and detecting features in massive data sets and AI is meticulous, things that humans are not very good at.
AI can filter out signal “noise” that has been properly flagged to it and it can crunch more data than a legion of interns using Excel ever could.
More useful though is that AI has gained traction when it comes to non-player characters in computer games — the “bad’ guys have gotten better at attacking human characters in the gaming environment and these “bots” have applications in approximating how a human trader would act in particular circumstances.
But while AI may be some ways away from approximating human behavior in financial markets, in digital asset markets, it has been making significant inroads.
For starters, digital asset markets are far more controlled environments than what many imagine them to be.
Because digital asset markets are highly manipulated and are populated with market-making (wash trading) bots which amp up volume artificially, the “genuine” human led trading activity is relatively easy to discern.
Using AI, it then becomes possible to filter out the signal noise and zoom in on human trading behavior.
But because digital asset markets trade 24/7, bots still lead in terms of trading volume, regardless of whether or not these bots are manipulated by humans or purely autonomous, behaving specifically according to preset parameters.
What has resulted then is that AI programs designed to trade digital asset markets have a somewhat less harsh learning environment than their financial market counterparts.
Part of that has to do with the relative immaturity of digital asset markets and the other part has to do with the lack of mainstream market participation.
For now at least, it’s a lot easier to train AI in the digital asset markets than it is in the financial markets, simply because there is far less competition.
But that doesn’t mean that AI, even in the digital asset markets has achieved anything even remotely close to clairvoyance — not by a longshot.
Ultimately AI is all about statistics.
Even the most highly attuned AI program, given the current state of art, can only achieve a probabilistic prediction of outcomes — anything beyond that is in the realm of science fiction.
Image Credit: Star Wars “A New Hope” Lucasfilm, ©1977
Patrick is an innovative entrepreneur and a lawyer passionate about cryptocurrencies and the business world. He is the CEO of Novum Global Technologies, a cryptocurrency quantitative trading firm. He understands the business concerns of founders and business people helping them to utilise the legal framework to structure their companies to take advantage of emerging technologies such as the blockchain in order to reach greater heights. His passion for travel, marketing and brand building has led him across careers and continents. He read law at the National University of Singapore and graduated with Honors in the Upper Division and joined one of Singapore’s top law firms, Allen & Gledhill where he was called to the Singapore Bar as an Advocate & Solicitor in 2005. He created Purer Skin, a skincare and inner beauty company which melds the traditional wisdom of ancient Asian ingredients such as Bird's Nest with modern technology. In 2010, his partner and himself successfully raised $589,000 from the National Research Foundation of Singapore under the Prime Minister’s Office. He has played a key role in the growth of Purer Skin from 11 retail points in Singapore to over 755 retail points in Singapore and 2 overseas in less than a year. He taught himself graphic design, coding, website design and video editing to create the Purer Skin brand and finished his training at a leading Digital Media Company.