YWR: AI and the importance of 'Why?'
Disclosure: Personal views only. Not investment recommendations.
I increasingly see a lot of investment strategies which claim to use ‘AI’ in their process. For example, an ‘AI’ driven Smart Beta strategy that is always long the S&P 500, or Bitcoin, but the AI knows exactly when to reduce exposure a little bit right before a sell-off and then knows to buy it back on the lows, etc. Supposedly. Then there is of course the growing use of AI in consumer applications from Netflix, to Spotify, Amazon, airplane ticket pricing and medical diagnosis.
My problem with AI, especially when it comes to investment strategies is that clients being pitched these strategies are afraid to ask how it works. They just nod their heads and go along with it. Probably because it sounds complicated and they don’t want to get into a statistics conversation that quickly goes over their heads. And if there really was a conversation the client would eventually find, the hedge fund that built the AI doesn’t really know how or why it works either. After lots of pulling on the string the answer would eventually be some version of ‘it just does’.
When Netflix knows I’d probably like to watch Ozark because I watched Narcos, it is based on conditional probabilities and Bayes Theorem. What is the probability of A given B? Or, what is the probability of C if both A and B are true? Netflix calculates the probability of Erik liking Ozark P(Ozark) given B(Erik likes Narcos). This is what people refer to when they say ‘Bayesian Inference’. As we know Netflix is always updating and refining these probabilities with its incoming stream of data and this is why it is an attractive business model. The data becomes a moat. The more data the better the AI. This scales up many times in computer vision. What is the probability this is a Japanese Spaniel? The computer needs to see lots of Japanese Spaniels and calculate the probability of each image of Spaniels given different colors, positions and backgrounds. The more spaniels it’s seen the better it gets.
Then the next level is Tesla and self-driving cars. Tesla takes all the data from its fleet to train a car to recognise lane markings, traffic signs, cars and bicycles. It’s magical how it all works, but when I was watching the Tesla Autonomy Day presentation in 2019 there was a key part that struck me. Tesla’s head of computer vision was explaining situations where the AI gets confused. One example was a Tesla car pulling up behind a car with a bike on a bike rack. The Tesla AI wanted to treat this as two different objects. There was a car and then also three feet off the ground in the middle of the air was a bike. The Tesla engineers then had to teach the computer that in this situation it is a bike attached to a car and is one object. It is called a ‘corner situation’, ie rare, and the benefit of Tesla’s large dataset is they can find all of these situations, teach the AI what to do and going forward all Tesla cars know how to respond.
Other situations that came up for Tesla are large trucks that fill all four cameras on one side of the car. The engineers needed to teach the car this is one large truck, not four separate trucks. Or, that if a Tesla car is stopped at a stop sign at an intersection and there are pedestrians crossing the street on the other side of an intersection, and while waiting at the stop sign, a large delivery van drives by, and the computer can no longer see the pedestrians, that they are still probably there walking across the street.
But the example of the bike illustrates a key aspect of how AI works. You and I know what a bike does and how it works. We have ridden bikes. We know that if they aren’t moving forward they fall over. We know that sometimes it is nice to put them on the back of your car and drive them to a scenic place for a ride. So when we see a bike 3 feet off the ground, not in motion, on the back of a car, our brain is associating a lot of situational data and also understanding of the broader environment including rules of gravity to understand what is going on and to focus on the car and to ignore the bike.
“But, Daddy Why?”
One of the things about kids is they are always asking ‘why?’ Why this, why that all day long. They see something in nature and want to understand the rules governing that object. Isaac Newton gets hit on the head with an apple and wants to know why the apple falls straight down, which makes him relate this to the sun and then finally to wonder if maybe apples fall because the Earth if actually a sphere circling the sun and actually all mass has an attractive pull and there is a thing called gravity. AI doesn’t do that it. It doesn’t care why. It just observes that statically apples fall from trees with a high probability. But going back to kids. They go to school for 12 years and learn the why of everything. They learn about monarchs and why we have a constitution. They learn about photosynthesis and why plants are green. They learn about gravity and physics. Kids will say everything they are learning is useless, but actually they are building a working model of how objects in the world work, and the reasoning and forces behind them. So a human knows there is a force called gravity, how it works and why. So even if a human has never seen a piano sitting in a tree, it knows what is going to happen if the branch breaks. AI has never seen a piano in a tree so until the piano falls and it is in the dataset it is clueless what is going to happen next. It makes you step back and appreciate the value of everything we learned in school and the reasoning behind the curriculum in elementary education.
Maybe in lots of things if there is enough data it doesn’t matter if the computer knows why something is happening or not. The same way a kid can speak English even if they don't know they the underlying grammar and what is a subjunctive tense. But with investment strategies I have a problem. Markets are reflexive in a way most of the physical world is not. What I don’t like is that statistically driven AI investment strategies are always backward looking. As Chris Cole at Artemis says it’s like teaching a car to drive across the Great Plains from Iowa to Denver, but then the car gets to Denver and has never seen a mountain before. AI strategies are also not aware of the broader context in which they are operating. The AI strategy doesn’t step back and go, “Hey Erik, Good morning. How are you doing? By the way, I noticed everyone else in the market is using the same data as me, and we are all optimising mean/variance in the same way and all buying the same stocks, and it is working, but that is because the amount of money being allocated to AI strategies is growing and so if it is a self-reinforcing cycle. And so, even though we are doing well. I suggest we take some profits and I should actually reprogram myself to now start buying stocks that statistically have been awful and none of the other AI programs are looking at like European Banks.” But no. They don’t do that. So for me that is a problem with AI strategies and it goes back to asking ‘why’? The AI never asks or digs into why the strategy is working, so it never sees the upcoming cliff it is about to drive over.
Another example in the real world is GPS navigation. Maybe this is just me, but if I suddenly get stuck in a traffic jam on a five lane freeway and Siri suggests I exit the freeway and take a 2 lane country road to circumnavigate the traffic, I increasingly don’t take that advice and stay on the freeway. What I’ve noticed is we all have the same GPS and all the drivers are getting the same advice and so what is going to happen? Several hundred cars are going to exit the freeway onto to a 2 lane road with traffic lights and very quickly that idyllic country road is going to be worse than the freeway. Meanwhile, after some of the cars exit, the freeway will be better and start to move again. So best to sit tight and wait for everyone to blow themselves up on the backroads. So ‘No Siri, thank you for your advice, but let’s just sit tight. Watch what happens.’
An increasing problem in AI driven business models is discrimination. This is coming up with business models that use AI for credit approval, insurance and job applicant CV screening. AI is superhuman at analysing thousands of data sets and finding patterns our human brain cannot. But there is no way to retrospectively ask the AI why it rejected a seemingly good mortgage application. Maybe the AI program just noticed the applicant had increased the amount they were gifting to their kids in their recent tax returns and also that they had increased their life insurance coverage and somehow figured out that this was a pattern they had seen before and this applicant had an increased % of committing suicide in the future. The problem is AI doesn’t know why it made the decision. It just sees statistical patterns. The legal problem for companies is if AI is unknowingly discriminating people based on gender or race. This is why there is a big push to retrospectively explain AI decisions, but it is shaky ground, because really there is no way to know. There are no rules behind AI decisions. There is no coding line to point to on why the application is denied. The computer is just teaching itself statistical probabiliies based on the data.
Cancer and COVID Masks
I came across a great example of AI pattern recognition and for me it shows how this can work well for investing. A friend of mine has an online luxury goods company and spends a lot of money advertising on social media. Advertising on the internet is difficult though. It’s hard to know where to direct your spending. Which sites, which key words, what kind of customer profile, how much to bid, etc. People think online retail is easy, it’s actually much more complex. My friend was approached by another entrepreneur who had developed an AI system for identifying customers to target and where to spend advertising money on social media. He promised he could improve my friend’s ROAS (return on advertising spend). As an example, this entrepreneur said he had been trying to sell COVID masks over the internet and had used his own AI program to decide where and how to advertise. The AI did its thing and spent a disproportionate amount of money advertising on Google Search to anyone searching about ‘cancer’. If you were searching about cancer you would be shown an advertisement for a COVID mask, and it turns out this worked very well.
My friend and I thought it was an interesting discovery. Probably, people with cancer, reading online articles about cancer, are in a fragile health condition and are statistically more likely to react to a COVID mask ad, but none of us would have thought of that naturally. We needed the AI to find that connection, but then we had to figure out what was going on. The AI doesn’t know what cancer is or why advertising to cancer patient works. But then it naturally made us wonder what other diseases are people reading about where they would also buy a COVID mask? To me it is this combination of AI pattern recognition with human curiosity that could be powerful. As investors we can have powerful AI programs recommending trades and strategies, but the key is to also dig in and try to find out why this is working. What is this telling us about the market? Is there a new industry trend? Is there a style rotation happening? Is this just all the AI doing the same thing?
So when it comes to AI and investing, we don’t just turn off our brain and let the AI invest, instead we work with the AI to maybe do the trades, but also learn what insights it is telling us about the market. What pattern is it finding? Then we do what we humans are good at, and what we’ve done since we are kids and start asking the ‘why?’ of it all. This is the way.
It’s the weekend and I’ll probably get a ride in, but I have to do a 15 min graded video presentation for the Econ course I am taking.
Have a good weekend!