The first thing coffee did was bring people together. In the half century following coffee’s introduction in England, coffeehouses had proliferated as a hub where people met, and caffeine and information flowed. One of these, a coffeehouse Edward Lloyd opened in 1687, became something much more. Lloyd’s was near London’s docks and became a hangout for ship workers. Lloyd understood the full nature of his business and with it the demand for information. This generated the unlikely complement, “Lloyd’s List,” a pamphlet that compiled shipping news—who was coming and going and sea conditions. Coffee may have been imported to England from the East, but the information was global.
What was all this information for? While idle gossip is part of the story, the other part was advantage. Gambling was afoot, and there was a healthy betting market. And not just betting on which bugs would be faster at climbing a wall, but whether, say, Admiral John Byng would be executed for incompetence following a French naval battle. That turned out to be a good bet, but what about whether a ship would return before a certain time? It would be helpful to have some knowledge of the ship’s route, the weather along the way, and what the ship might have to procure to make the return trip more lucrative. Making a better prediction would help you beat the odds.
While gambling can be motivated by fun or addictive dopamine, as a sustainable business proposition it needs something more. The innovation came in noticing that offering bets was a way for risk mitigation rather than risk exposure. In shipping, a lot of that was resting on ship performance. For starters, the capital investment was huge, so if a ship itself (let alone its cargo) did not return, those financing the trip could face large losses. But even delays could create lost opportunities and impact on other markets. If your business relied on imports, then a delay of a shipment could mean ruin. In other words, businesses were gambling all the time. What they wanted was insurance to turn a gamble into certainty.
Let’s reflect here on decision-makers’ desire for certainty. The traditional go-to explanation in economics is risk aversion. A decision-maker is risk averse if they chose not to take a fair bet. For instance, if you were offered $60 when a dice roll turned up a three and nothing otherwise, then you are risk averse if you would prefer to be given $9 (or something less than $10) instead. This logic also works in reverse. Suppose you currently face a risk that if the dice roll a three, you would have to pay out $60. Then you are risk averse if you would pay $11 (or something more than $10) to be freed of that potential loss. In one situation, you decline a bet with an expected return of $10 for a certain payment of something less. In the other situation, you will avoid a bet with an expected loss of $10, even if it costs you more than that. In one case, you pay to decline risk, and in the other, you pay to have risk taken away. Thus, if you are risk averse, you avoid gambling and you buy insurance.1
In this chapter, we consider how decision-makers deal with uncertainty when they have good reasons to minimize their exposure to uncertainty. From this, we are able to more precisely express what better prediction is doing for people. We begin with the most familiar way of reducing exposure to risk—taking out insurance. We point out that by doing so, not only can you do better if you are risk averse, but you can also reduce the need to gather information for the purpose of making a decision. Of course, insurance is perhaps the clearest instance of a more general way of dealing with risk exposure—hedging your bets. Our interest here is in how prediction itself can provide information that allows for better decision-making by managing risk and so obviates the need for insurance to limit risk exposure.
Insurance Coverage
The desire for certainty created the insurance industry, as exemplified by the activities in Lloyd’s coffeehouse, which today goes under another name, Lloyd’s of London, sans coffee and now a centuries-old insurance operation. The coffeehouse information became invaluable in something new—calculating the odds of various events that businesses wanted to place bets on for the purposes of insurance. That information was critical because those who took those bets—underwriters—needed to be paid for the privilege. They had to earn enough from premiums across all their clients to cover potential losses, with a bit extra for their part in the client’s risk management. Better prediction fueled appropriately priced bets and allowed what might be variable returns for an individual business to be turned into steady returns for a group of them.
Uncertainty pervades decision-making. And most decision-making, in business and elsewhere, takes place without an explicit insurance market. To be sure, maritime insurance exists to this day. It is likely considerably cheaper as shipping became safer but also somewhat more exotic as new products such as ransom insurance (in case your ship is taken by pirates) emerged and became predictable enough to be priced. But elsewhere, for a variety of reasons, people make decisions without the ability to make a side bet to take the edge off risk.
When they are potentially exposed to risk, what decisions do people make? If the decision itself does nothing to change their risk exposure, then the risk will not really matter. But it is precisely because the actions that might be chosen might have different risk consequences, that risk exposure matters. To see this, consider a farmer who is considering how much land they might cultivate for a maize crop. The yield of such a crop heavily depends on the forthcoming rainfall. High rainfall means a high yield; low rainfall can mean a low or no yield. If we add to this situation, as might arise in many developing countries, that the crop yield directly flows into what a farmer’s family might be able to consume, the rainfall risk can be considerable. Maybe it would be better to save money or not go into debt than expose themselves to that risk.
As rainfall is a measurable event that is also not in the farmers’ control, an insurance product based on rainfall outcomes might change decisions. That product would pay out if there was low rainfall, so a farmer could pay off debt or have income for their family’s consumption. In effect, with insurance, a farmer is betting on the rain. But unlike gambling, they are betting against rather than on good news. On the one hand, by cultivating more, they are betting on high rain, and on the other, by buying insurance, they are betting on low rain. Do both and it is a wash; they are making no bets at all. If risk exposure is what is holding cultivation back, then adding insurance should change that decision. If risk exposure is not the constraint, then insurance won’t do anything.
Fortunately, thanks to economics researchers, we do not need to theorize about this possibility.2 Those researchers conducted a multiyear study whereby they gave farmers in Ghana cash grants, rainfall insurance, both cash and insurance, or neither of the two to measure their impacts on cultivation choices. The researchers found that cash grants were not effective in improving agricultural investment decisions, so it wasn’t just that the money to plant a crop was hard to come by. Instead, it was all about risk exposure. Those who had access to the insurance product, on average, invested between 10 and 15 percent more in cultivation, even though they rarely insured more than 60 percent of their land.3
What this tells us is that insurance can change decisions toward what might otherwise be risky options. But it does more than that. If it were possible for Ghanaian farmers to spend time and money to forecast rainfall for the coming year, they could also de-risk their decisions by forgoing cultivation in years when little rainfall is predicted. In other words, insurance obviates the need for prediction as an input to decision-making. In the Ghanaian case, those forecasts do not exist, at least not with sufficient confidence to be able to mitigate their risk exposure. But as we will discuss shortly, when insurance is unavailable, decision-makers actively seek out predictions so that they can use them to mitigate the costs of risk and uncertainty.
Predictions Mitigate Risk
As we talk to business leaders about AI and how, as a technological advance, it reduces the cost of predictions and thereby provides more opportunity to employ predictions, our go-to example is often the weather. The weather is something that we know has an impact on our day-to-day decisions. Whether to pack an umbrella, wear a coat, take a different route; we could go on. The weather is variable, and it matters for our decisions. We don’t need to convince anyone that better weather forecasts are valuable.
Our daily decisions are one thing, but there are situations where weather risk is potentially extreme and weather forecasts, even imperfect ones, are invaluable. This is perhaps no more apparent than at wartime. Numerous battles in history had outcomes that depended on weather. The most famous was the D-Day invasion of Normandy, which required the avoidance of potential storms on the English Channel. Thanks to weather forecasts, the right call was made.
All sides in World War II knew about weather risk. But by dint of geography, there was an asymmetry in each side’s ability to forecast the weather. Storms and other weather systems tend to move east across the Atlantic. The United States could measure those systems and thereby provide a more accurate forecast of weather on the other side of the Atlantic. By contrast, the Germans had to go to great lengths to obtain the relevant data. They could send ships, but the ships couldn’t stay in one place long because the Allies were hunting them. Thus, in September 1943, a German U-boat took a team across the Atlantic and landed in Canada to build a weather station.4
This was no mean feat. There was the issue of where to put it. It had to be remote, but places are remote for a reason. The U-boat captain landed in North Labrador. Then the Germans had to automate the weather station, including for power; they had brought along ten very heavy batteries. They then set up the station and tested its transmissions. They painted it with the name Canadian Meteor Service in the hopes that no one would touch it, even if it was discovered. It took the Germans a little over a day to install the station before they left to return to Europe.
That weather information may have helped Germany regain some of that lost weather forecasting deficit. But for reasons that are still unclear, the transmissions were blocked a few weeks later. The Germans sent another mission to Canada in 1944, but the U-boat didn’t make it. The weather station’s existence was finally discovered by accident in 1980.
Better weather predictions would have allowed German leaders to move military activities around to avoid weather events. However, the predictions would also have allowed them to better forecast Allied moves. After all, the Allies had the forecasts. On D-Day, the inferior German forecasts gave them excessive confidence that no attack would take place, and they relaxed accordingly. General Eisenhower would eventually credit the Allies’ superior meteorologists with the European victory.
The point of this story is not to support the rather obvious claim that predictions are valuable. Instead, it is to demonstrate precisely when they are valuable. Predictions can turn an otherwise risky option into a less risky one. It is risky to plan an offensive during a certain window of time when you don’t know what the weather will be like and can’t cancel the operation once you find out. But if you have a forecast, you can plan around those events and choose your timing in a less risky manner.
How Confident Are You?
Does this suggest that having a prediction leads to better decision-making? That is not quite the case. The quality of the prediction matters. During World War II, Ken Arrow, later a Nobel laureate in economics, worked as a weather forecaster for the US Navy. His team of statisticians had the task of forecasting the weather a month ahead. They discovered that those forecasts were no better than random guessing—in other words, they were not forecasts at all. The team suggested it be tasked with something else. The answer from higher up was no. According to Arrow, the reply read: “The Commanding General is well aware the forecasts are no good. However, he needs them for planning purposes.” Orders are orders.
This story, crazy though it is, does highlight something that simple discussions of prediction can obscure. Not all predictions are equal. This is most definitely true of weather predictions. For decades in the eighteenth century, farmers across the United States purchased the Farmers’ Almanac, which provided weather predictions a season or a year out, despite such predictions being no better than a groundhog. However, scientific weather predictions have improved markedly. First, this was done by gathering data the way the German military was trying to do. Then, in modern times, atmospheric scientists have used computers to build models that, when combined with that data, predict weather in a particular location to an astonishing extent. A five-day forecast is now as accurate as a one-day forecast a decade and a half earlier.
But how should we use that five-day forecast? Five-day forecasts have errors, and at times, the weather is less predictable than at other times. Weather forecasters knew about this unpredictability because they ran forecasts through multiple models. When they agreed, the forecast was much more reliable than when they did not agree. That means that if your decision relies on the forecast five or six days out, you want to know its reliability. According to DJ Patil, who was the first US chief data scientist, weather services should vary how far out they actually provide a forecast based on that reliability.5
No prediction is 100 percent accurate. And in the realm of AI where predictions are becoming better over time, one of the things we mean by “better” is that they are becoming more reliable. The critical question for a decision-maker is when is a prediction reliable enough. For instance, one of the major costs of modern farming is for fertilizer. But if there is rain soon after fertilizer is put down, that expenditure is wasted. So, what causes you to flip a decision based on a forecast of no rain? Suppose that you would take the risk if the forecast of rain is 5 percent over the next week. However, when you receive a forecast from a weather service that is below that threshold (say, 4 percent), is that enough? If the reliability of the forecast is 95 percent, then the probability of rain is potentially greater than 4 percent. What’s worse is that even reliability isn’t enough. If a forecast is accurate 95 percent of the time, it might mean that it predicted rain but there was no rain 5 percent of the time, or that it predicted no rain but there was rain 5 percent of the time. The point here is that when forecasts are imperfect, it becomes important to know what is going on under the hood. As predictions improve initially, by relying on them, you might not perform much better or even worse than you currently are. But eventually, performance will rise as reliability rises.
On-Time Prediction
Another dimension by which we mean that a prediction is better is with regard to speed. It is a fairly commonsense notion that a prediction made after the event it is trying to predict has occurred is not a valuable prediction. What is interesting is that there are many areas where we know how to predict an event, but we cannot generate the prediction in time. Again, the advances in weather prediction illustrate this nicely.
In the mid-1800s, telegraph lines were being strung across the United States. People sent messages along those lines. But the weather sent its own message, as the telegraph didn’t work in the rain. That meant that an East Coast operator would know if a storm was on the way because the lines in Ohio were down the previous evening. For the first time, information required for weather prediction could outrun the weather. As Andrew Blum writes, “[O]nce the news could travel faster than the winds, then the wind would no longer come as a surprise.”6
In the nineteenth century, the constraint on weather prediction was data—just as it was for the Germans in World War II. But communication allowed the data to get to where it was needed. In the twentieth century, the ability to generate forecasts using a model or simulation of global weather systems was constrained by the sheer complexity of the calculations. With older computers, an accurate prediction of the weather twenty-four hours in advance was possible, but it would take more than twenty-four hours to calculate. The advances in computer technologies—especially supercomputers—relaxed that constraint.
More recently, providers of weather forecasts faced a new constraint. Mobile phones and apps changed how consumers used weather forecasts. They were no longer content to rely on weather forecasts generated the night before. Instead, as conditions changed, the minute-by-minute forecast could be updated. Consumers checked weather on the internet and apps on their mobile phones and wanted predictions for the next hour or two that required the latest data and calculations for accuracy. The problem was the entire system of forecasting was not set up to operate at that speed.
This was the insight of Peter Neilley who worked for the Weather Channel (now called The Weather Company). The constraint was now human. To ensure the accuracy of local forecasts, no forecast could be produced without the signoff of a human forecaster. That was fine for day-out forecasts but would be a problem for hourly updates. Neilley realized that automation was required. In the end, there was a reorganization. Under normal weather conditions, an automated system was deployed, and forecasts could be made without human signoff. For extreme predictions—such as storms or tornadoes—a human had to be part of the process. Forecasts moved from being scheduled to being on demand. The expectation was that there would be some loss in accuracy. But the market conditions suggested that the trade-off toward speed was worth it.
When predictions can be provided quickly and updated continuously as new data becomes available, this changes how decision-makers use that information. Consumers understand that weather forecasts are more accurate the closer they are to the day in question. There is a cone of uncertainty for any future prediction whereby forecasts are less reliable the earlier they are made. This means that decision-makers face a new option: when to decide.
When you should decide can be a complicated matter. Often decisions have their own time. For instance, you are not going to decide whether to carry an umbrella after you have left the house. But in other situations, there may be value in waiting. You might be planning an outdoor adventure a week in advance, and a weather forecast can tell you if that is a good idea. But if you can wait, there is what economists term “option value” in not committing to being outdoors until as late as possible because you might be able to base that decision on a more reliable weather prediction.
The availability of more timely predictions means that there is value to taking more time before you decide to do something. That means changing our habits and also potentially our workflows to take advantage of those real options. In this regard, better prediction does not necessarily mean that our decisions become easier to make, but instead, that there is value in making our decisions somewhat harder in terms of what options we throw into the mix.7
Better Prediction Reduces the Need to Manage Risk
There are two broad ways to manage risk.8 The first, which we have already discussed, is insurance. That is, you take actions that reduce the costs associated with bad outcomes. This involves improving the downside outcome you might receive from taking a risk. For example, you can fertilize your crops without bearing too much risk if you also purchase insurance that provides a payout in the event of rain. The second way is by protection. That is, you take actions that reduce the probability of a bad outcome. This is another way of saying you take fewer risks, perhaps by planting crops that are more weather resistant.
We can see this impact by looking at how the adoption of AI prediction had an impact on Air Canada’s cargo business. Air Canada knew there was room for improvement in how it handled cargo. On capacity-constrained routes, 20 percent of shipments were no shows (shipments that were planned but didn’t happen), and overall 45 percent of capacity was unfilled.9 The planes were flying, and they could carry more. It was far more challenging to predict cargo than, say, passengers. There were last-minute passengers, but that number was proportionately small compared to last-minute cargo. And unlike people, the booked weight of cargo could change appreciably at that last minute. There was a ton (pun intended) of multifaceted uncertainty. That made it a good problem for AI prediction to solve.
There were several things in Air Canada’s favor in deploying AI in cargo booking. First, the waste and, hence, opportunity was there for all to see and it was measurable. Second, to predict cargo demand, Air Canada had a wealth of data. It could use this to assign probabilities to different customers that the loads that were booked would actually fly. Finally, the entire problem was contained internally within Air Canada. Not surprisingly, it was able to produce a prediction machine that promised to drop unused capacity by an average of a quarter. The only wrinkles came in implementation. Handlers had to deal with more cargo and pack it efficiently. They didn’t have the knowledge to do that, so significant retraining had to occur. But that was a one-off investment, the sort of thing you expect when having more information and a desire to use it.
Air Canada’s model of allowing for underutilized capacity when it did not have AI prediction was akin to taking out insurance. It left a buffer to ensure it could deal with surges in demand for cargo without causing delays. This enabled Air Canada to offer a service that allowed its customers to change their minds at the last minute. As an alternative, it could have chosen to target customers who were not so seemingly fickle; for example, customers who themselves had regular, stable shipping needs. This would allow it to book long-term contracts for haulage where the bad outcome of overbooking could be avoided. The fact Air Canada didn’t do this tells us something about the value of those customers. As it turns out, those with stable needs do not necessarily pay the most for air shipping. But the point is that had Air Canada targeted customers with stable needs, it would have had no waste. This is a risk strategy of protection.
Interestingly, in the absence of prediction, when you use insurance to minimize waste, the waste is visible. By contrast, with protection, the waste may be harder to see. Thus, even though the underlying uncertainty is the same, it is not hard to imagine situations in which businesses choose to adopt AI when waste is visible and not adopt it when it is hidden.
KEY POINTS
- The trade-off between gambling (taking your chances) and insurance depends on the quality of information available to the decision-maker before choosing an action. As prediction improves, decision-makers will switch from insurance to gambling in terms of managing risk.
- Risk exposure matters because different actions might have different risk consequences. Consider a farmer who is considering how much land they might cultivate for a maize crop. The yield of such a crop heavily depends on the forthcoming rainfall. High rainfall means a high yield; low rainfall can mean a low or no yield. As rainfall is a measurable event that is also not in the farmers’ control, an insurance product based on rainfall outcomes might change decisions. That product would pay out if there was low rainfall. By cultivating more, they are betting on high rain, and by buying insurance, they are betting on low rain. Do both and they are making no bets at all. If risk exposure is holding cultivation back, then adding insurance should change that decision. If risk exposure is not the constraint, then insurance won’t do anything. This tells us that insurance can change decisions toward what might otherwise be risky options. But so can enhanced prediction. If it were possible for farmers to better forecast rainfall for the coming year, then they could de-risk their decisions by forgoing cultivation in years when little rainfall is predicted. In other words, enhanced prediction, like insurance, can turn an otherwise risky option into a less risky one.
- As an alternative to gambling, there are two broad ways to manage risk: (1) insurance, and (2) protection. For example, airlines face uncertainty with respect to the demand for cargo—often as much as half a plane’s cargo capacity is booked at the last minute. So, airlines face a risk of overpricing (and losing customers) with the hopes of attracting last-minute rush orders, or underpricing and losing revenue from higher-price last-minute customers. An insurance-type approach is the one often taken, which is to overprice and underbook, leaving a buffer to benefit from surges in demand. A protection-type approach is to enter into long-term fixed-price contracts with customers who have regular shipping needs, even though this means charging lower prices overall because regular customers refuse to pay last-minute booking prices. The opportunity for improvement is obvious in the insurance case because it’s easy to see planes flying with unused capacity. The opportunity is more difficult to spot with the protection approach when planes are flying full and only the careful observer notices that the long-term fixed-price contracts are priced significantly lower than just-in-time shipping. Enhanced prediction would increase the returns to gambling and require less insurance because more accurate forecasting facilitates better pricing for the spot market, increasing capacity used and reducing reliance on long-term fixed-price contracts.