Avi Goldfarb is the Rotman Chair in Artificial Intelligence and Healthcare and a Professor of Marketing at the Rotman School of Management, University of Toronto. He is also Chief Data Scientist at the Creative Destruction Lab, Senior Editor at Marketing Science, a Research Associate at the National Bureau of Economic Research, and President of Goldfarb Analytics Corporation.
In his new book, “Prediction Machines,” Goldfarb explains how artificial intelligence (AI) will affect business, public policy, and society in terms that work for decision makers in virtually all fields. Perhaps his most important role is coming up this October in Boston, when he will be the featured speaker at the FP&A Luncheon at AFP 2019.
Goldfarb recently joined us on the AFP Conversations podcast, and we discussed using AI to make predictions in all areas of life and business. Here are five takeaways from that interview.
The AI age is distinct from, and a successor to, the internet age. “There's something different about what the current generation of artificial intelligence, which is primarily machine learning, can do, relative to what we saw with the internet. So when I think about the internet, I think about it as making communication easier, and making copying replication of digital files easier, and making search easier, and those were transformative to the economy. When I think about the current generation of AI, it's prediction technology. It's fundamentally computational stats. And so that makes it much easier to fill in missing information. A lot of what we do as managers or as humans generally, involves filling in missing information. And so a technology that can help us do that is transformative.”
Economics 101: AI turns “prediction” into a low-cost item, which leads us to use more of it. “When you think about artificial intelligence in the form of machine learning, you could think about it as making prediction cheaper. When something becomes cheap, it also affects other inputs. When coffee becomes cheap, the first result is, we buy more coffee, we know that. The second result is we buy less tea. Tea and coffee are substitutes, if coffee is cheap, you'll buy coffee instead of tea. And then thinking about prediction, machine prediction and human prediction are substitutes. So, those aspects of your job that are prediction are increasingly going to be done by machine.
“So the big question is what are the complements to predictions? What's going to become more valuable as prediction becomes cheap? And in order to understand that, you need to recognize that predictions aren't useful in and of themselves, they're only useful because they help you make better decisions. And so the other aspects of decision-making are the things that become more valuable.”
AI and machine learning can fill in (“predict”) the missing information in financial models. “Some things that we didn't use to think of as data that we could put into our models, now can be seen as data because we now have a tool for filling in missing information. We now have a tool for understanding a map or a satellite image and turning it into information that can be processed and used to predict outcomes in financial markets.
“So, your key role for humans is creative identification of new data sources. Another potential use is, some things that we used to think of as unpredictable around the world, that we didn't have good financial models for, with these techniques potentially if they're high frequency it could be ,used for prediction.”
Humans have three key roles to play relative to machines in AI. First, “‘Automation’ means a human figures out, 'Here's the thing I want to predict,' and identifies a way to, for example, buy and sell under that prediction model. And then a machine takes the actions of posting prices, buying and selling, or putting together portfolios et cetera.
“The second is ‘human in the loop,’ where a machine might make recommendations, but no action is taken until a human looks at those recommendations and assesses whether they make sense. And for high stakes decision-making, especially for things that don't require milliseconds for transactions, human in the loop is a useful way to think about it.
“Then there's something in the middle which is ‘human on call.’ And human on call is the idea that usually you can automate, but when something is unusual, say an unusually large transaction or something the system hasn't seen before or something where the machine's predictions are uncertain like a high confidence interval, high variance, then you can call on a human to make the final action. So usually you'd have the system automated, but every once in a while you bring a human in.”
The decision to deliver tightly bound historical trends or varying degrees of serendipity is a human-made design element. “Think about a recommendation algorithm. So one version of a recommendation algorithm is people who bought this item also bought that one. So the classic example of that is, 15 years ago, Amazon was rolling out their recommendation systems, and almost everybody got recommended the Harry Potter books. Why? Because no matter what you bought, it's very likely you had also bought the Harry Potter books, because so many people did. And so that's a recommendation system design that doesn't lead to much serendipity. That leads to essentially a superstar effect, where everybody gets recommended the same super popular things.
“Over time Amazon and others tweaked their recommendation systems from that to people who bought this disproportionately bought. And that leads to unique recommendations that are distinct to you with subsets. So if you had bought the book, “Prediction Machines” for example, a long time ago you would have been recommended, Harry Potter. But increasingly, you're going to be recommended other similar books on the impact of artificial intelligence.”