Source – techcrunch.com
From hedge funds to venture capital firms, everyone in finance has some idea about how data and quantitative analysis will reshape their industry. Firms likeSignal Fire track engineers as they move from company to company to draw attention to growing startups. And funds like Numerai and Quantopian are putting faith in quants to determine optimal trading strategies.
Bridgewater Associates, one of the world’s most robust and reliable money making machines, is going as far as to attempt to automate its internal management processes to ensure the longevity of its $150 billion under management. But unlike most other approaches of applying AI to moneymaking, Bridgewater’s tactic isn’t about anomaly detection, it’s about mechanization.
It was about the people before AI
To understand Bridgewater, you have to understand Ray Dalio. To Dalio, broken frameworks and excess emotion are the enemy. Success comes from a curated set of rules he refers to as Principles in his book of the same title.
The field of behavioral economics is dedicated to studying the myriad of ways that psychology and neuroscience influence decision making. Traditional economics makes basic assumptions about human rationality but research in behavioral economics has shown us that people tend to do very strange things outside the paradigm of homo economicus.
There are hundreds of known cognitive biases — confirmation bias (we often only see info that validates our prior assumptions) , hyperbolic discounting (we’re really poor at valuing things with respect to time) and the bandwagon effect (we attach too much value to herd behavior).
Dalio says that rules help him to notice his biases and account for them. Whenever a conviction he has contrasts with what a computer model says it prompts reflection that can help to settle the dispute and lead to a better outcome.
The key is ensuring that you don’t overcompensate with your own emotions or do something just because a computer instructs you to. No number of algorithms can fully insulate a person from bias but they can aid in discipline and habit formation.
It will be about the people after AI
Decades ago, Dalio says he would write down his criteria for making a trade and then work to see if those criteria could be converted into an algorithm.
“When I think hard I can convert qualitative problems to quantitative problems,” Dalio noted. “I ordered a Cobb salad. If I could slow down I’d write down my criteria for a Cobb salad — qualitative judgment for liking a Cobb salad.”
This expert systems approach is antithetical to today’s conceptions of deep learning whereby a machine learning model is trained on massive quantities of data to produce a conclusion based on inductive reasoning.
“I don’t like the term machine learning because what I’m doing is not learning,” Dalio emphasized.
The distinction might seem petty, but it’s far from it. Many of the machine learning models in use today operate as black boxes — data enters and conclusions are spit out. If you want to ask what drove the model to come to those conclusions, you’d be unable to find any paper trail.
“If a machine comes up with an algorithm and you don’t have a deep understanding of the appropriate cause and effect relationship, than things get very dangerous,” Dalio explained. “If the future is different from the past, you’ll probably crash.”
Most data scientists today agree that it’s important to have some domain experience about the problem you’re trying to solve before you throw machine learning at it. This is important so that, say, weed plucking robots don’t get distracted with morning dew they never accounted for. Or in the case of Bridgewater, understanding is important to ensure that decisions aren’t made without an anchor to reality.
It’s for this reason that Dalio believes that the future of artificial intelligence will rely on humans. In his book, he notes that the day when a computer would be able to generally outperform a human without a human’s help is far away.
The key is understanding
Dalio believes the artificial intelligence of today breaks down into three categories — mimicking, data mining and expert systems. Mimicking refers to tasks that are easily replicable whereby understanding isn’t necessary. A characteristic of these problems is that they occur in worlds not subject to change.
Data mining opens things up to a broader set of problems. While not the term du jour, Dalio is referring to deep learning here where large quantities of information can be applied to solve specific problems.
The last approach, expert systems, is Dalio’s preference when he wants to ensure understanding. These algorithmic rules, derived from so called “experts,” are brittle and not widely applicable to today’s problems of object recognition and dialog systems but they can have value when deductive reasoning, rather than inductive reasoning, is required.
Deciding whether to fire someone, for example, is a deeply complex task for which data often cannot account. As Bobby Axelrod, channeling Dalio for a brief moment, in the TV show Billions put it, “There is a small group who can do the math. There is even a smaller group who can explain it. But those few who can do both… they become billionaires.”
If you try to force it, that is to say rely on the math when you don’t have understanding, you run the risk of flying too close to the sun. Dalio provided the example of merger arbitrage to explain just how these scenarios play out.
In merger arbitrage, a fund buys shares in the company being acquired and shorts the acquiring company. But the strategy can backfire if too many people bid up the price such that a savvy investor would be better off doing the opposite. (Check out What’s 2/3 of the Average in game theory literature for an interesting corollary here).
“There’s this assumption that big data sets are going to be the difference maker,” Dalio added. “But best idea is to have someone who can convert words into algorithms. If you can do that you’ll beat the giants.”
Dalio’s expert systems approach might sound dated, and it is, less we forget that deep learning and most of the rest of AI is also dated. We have no reason to assume that neural networks are the solution to recreating intelligence so for the time being we ought to accept that there are a number of techniques that can outperform each other in different contexts.
Make what you will of Dalio’s approach to investing, or to management or even to AI, but he is certainly right about one thing — pretending domain experience doesn’t matter in data science is a mistake. This is true no matter the machine learning approach, no matter the problem at hand.