Three weeks ago, I wrote a note titled “We’re Doing It Wrong“, which took to task the embedded anthropomorphic fallacies in how we use the massive computer processing capabilities now at our fingertips, what’s usually called artificial intelligence (AI) but is better named Big Compute. I won’t repeat all of that note, but here’s the skinny:
We think of markets as a clockwork machine, as an intricate collection of gears upon gears. We believe that if only we examine the clockwork closely enough, we can identify some hidden gear or unbeknownst gear movement that will let us predict the clockwork’s movement and make a lot of money.
Our model of markets is The Machine, and every Machine has a deterministic set of algorithms that create and drive it. Every Machine has an Answer.
This model – the market as machine – is an anthropomorphism.
If you use computers in your investment research process – and I know you do – I will bet you umpteen zillion dollars that you have those computers looking at structured historical data in an effort to find some repeating pattern. I will bet you do this rigorously and intentionally if you’re a quant. I will bet that you do this all the same, but non-rigorously and haphazardly if you’re not a quant.
Whether you realize it or not, you are using the market-as-machine model. You are looking for the Answer. Go on, you can admit it. You’re among friends here. I’m like Big Lou in the insurance ads … I’m one of you. It is embedded in our minds and in our businesses. Mine, too.
But here’s the thing.
If you use AI as just another input to that market-as-machine investment research process, you will get puzzling “results” that don’t help you very much. It will be just like using a telescope to get better measurements of the retrograde motion of Mars as it orbits around the Earth in your Ptolemaic model.
You will be disappointed by AI.
This is a note about how not to be disappointed by AI.
This is a note about how we can do it right.
And it starts with the game of Rock, Paper, Scissors.
I think we’re all familiar with this classic two-person game of strategic interaction, where on the count of three each player makes a hand sign for either Rock (a fist), Paper (a flat hand), or Scissors (two pointed fingers). Rock beats Scissors, Paper beats Rock, and Scissors beats Paper. It’s a game with (obviously) an enormous amount of luck to it, but human behavioral foibles and our poor ability to be truly random with our choices allow skilled players to do better than luck alone would predict. And as you might expect in a world where we are prepared to wager and compete on anything, aficionados of Rock, Paper, Scissors make up a highly devoted global sub-culture, complete with professional tournaments and dedicated YouTube channels and the like. It is a particularly powerful social phenomenon in Japan, where the game is called janken, and as likely as not any sort of minor disagreement or preference dispute among friends will be settled by a quick countdown and throw of hands.
So you can imagine the popular interest when a robotics lab at the University of Tokyo recently publicized its development of a machine that has a 100% win rate against human opponents in janken.
Fine, go ahead and build a computer to beat the reigning world champion of Go. But a computer that wins every janken game it plays? This is madness!
Luckily for the professional Rock, Paper, Scissors tour, however, this janken machine won’t be dominating tournaments anytime soon. Why not? Because it cheats.
The janken robot takes ultra high-speed images of the human player’s wrist and hand as it descends on the count of three, normalizes these images to account for different hand shapes and movement speeds, computes the final hand shape determined by the muscle and finger movements of the descending hand, and shapes its robotic hand into the winning shape before the human hand finishes its descent. The entire process takes about 20 milliseconds, with one of the toughest problems being slowing down enough so that the computer hand is formed at the same time as the human hand.
Twenty one-thousandths of a second.
Why do I love the cheating janken robot so much?
First, because it embodies what AI looks like. AI is not a giant brain. AI is massive computing power trained to “see” a little piece of the world in a way that the human brain simply cannot.
Second, because it embodies the best quote of one of the best investors of the 20th century,
Chancellor Palpatine George Soros.
I’m not predicting. I’m observing.
The janken robot is not a predictive machine, it is an observing machine. It is not evaluating the Rock, Paper, Scissors playing history of the human opponent in hopes of finding some predictive historical pattern of strategic choice. It is purely a forward looking machine with no need to analyze past plays of the game. There is no structured data to be processed here. There is no place to plug in a regression. It is all reaction, all the time.
The janken robot is not SOLVING for the future behavior of its opponent.
The janken robot is CALCULATING the future behavior of its opponent.
Doing it wrong means using our Big Compute superpowers to identify predictive algorithms by which we can solve for the market’s future.
Doing it right means using our Big Compute superpowers to visualize social behaviors by which we can calculate the market’s future.
Here’s a rough schematic for what a janken machine would look like in an investment context.
I first wrote this up in 2011, and I know that a lot of this design has been leapfrogged even in a schematic sense. It also pre-dates all of our work on the Narrative Machine, and I’ve intentionally left out some key parts in this schematic. Also, I know that a lot of this schematic will be gibberish for some readers no matter how much I explain it, and trivial for some readers no matter how little I explain it. So don’t @me. But here’s the basic gist.
Start at the top center where data flows into the machine. By messages I mean both structured data (prices, tickers, etc.) and unstructured data (publications, transcripts, news feeds, etc.). The replication server copies and shunts the message flow to three different AI areas: A (what hand shapes win this super-janken game?), B (what hand shapes will the other players make next?), and C (executive function to tell A and B how to work together).
Both AI A and AI B have a Complex Event Processing (CEP) module at their core, which is just a ten dollar phrase for an AI that’s been trained to evaluate the connections between disparate data from disparate sources in real-time or near real-time.
The Bot Queue in AI A is a set of static data models generated by AI C, the executive controller. Think of these as small probing trades that are constantly being made on different time frames to see what works. That is, we don’t have a constant rule set for what “wins” in markets like we do in Rock, Paper, Scissors. But we can observe what rules work NOW.
The Game Specification in AI B is a set of dynamic data models generated by AI C, primarily different iterations or flavors of the Common Knowledge Game. Think of this as the AI in the janken robot that recognizes different preliminary movements of the human wrist and hand and calculates where that hand will end up. Again, this can be designed to play out over whatever time scale you like, although the computing power required for an effective calculation gets larger and larger the farther out into the future you go.
Do we remember what we’ve done? Do we learn? Sure, and that’s where AI C, the executive controller, comes into play. But we’re remembering and learning with a totally different purpose than any other current AI implementation in financial services. This machine is relentlessly forward-looking and actor-oriented. We don’t care about the behaviors of the past. We don’t care about the statistical arbitrage of structured data and squeezing out a few dimes from ephemeral anomalies. We care about observing the unstructured conditions of the present and understanding the physics of what hand shapes will be made next. Because those micro behaviors of fear and greed can NEVER be arbed away.
Is this an expensive project? No, we’re not building a multi-billion dollar supercomputer to simulate nuclear weapon design. But yes, it’s a low to mid single-digit million dollar price tag to do this right.
We’d like to build this machine in 2019.
We’d like to find a strategic partner to build it with us.
This isn’t a Star Wars thing. This is a now thing.