The study of economics made a lot of progress in the late 19th and early 20th centuries by creating formal models of behavior that assumed rational actors. You can actually get a lot of mileage out of asking questions like "how would a selfish coal mine operator maximize the economic value of their finite natural resources?" or "how would someone who wants to smooth their consumption over a lifetime allocate between consumption and savings day-to-day, and when would it make sense for them to buy insurance despite a negative expected value?" And then economics got even more mileage out of relaxing those assumptions, and expecting people to be uncertain, inconsistent, impatient, misinformed, biased, etc.

One element of that was loss aversion, the claim that people are more upset by losing a given amount of money than they are gratified by making the same amount, such that a coin toss is utility-destructive. A tiny amount of loss aversion makes sense if there's diminishing marginal utility of money—if you're $5 away from being broke, a $10 gamble isn't worth it—but the initial claim was that loss aversion was a more general rule, and that in basically any circumstance it's an accurate description of how people operate.

This turns out not to be entirely true. There are contexts in which people are loss-averse, contexts in which they're actually more fixated on gains than losses, and cases where they're fairly neutral. And that makes sense because loss aversion is a reference dependent phenomenon. The existence of products like home insurance implies that people are willing to earn an expected 50 cents on the dollar in exchange for knowing exactly how much they'll lose and when. Meanwhile, negative-edge gambling is also wildly popular, in casinos, sports betting apps, and retail brokerages. Loss aversion makes the existence of gambling a daunting puzzle: the vast majority of games involve sub-50% odds of winning each round, with wins either paying out as much as losses cost (if the odds of a player winning are in the high 40s), or many multiples of that (if the player's odds of winning are low). There are specific strategies that can have a return profile that looks like an attempt to avoid losses, like a pass line bet in craps. And players can create a synthetic boring payoff function through a Martingale strategy where they continuously double their bet until they're back to even.

And loss aversion creates other puzzles. For example: why haven't real-time stock quotes and narrow spreads made every avid market-watcher miserable? If you watch individual stocks or your entire portfolio in real time, you see the P&L number constantly twitch up and down. If the twitches down were even slightly less pleasant than the improvements, it would make the trading day miserable. Loss aversion also implies that diversification makes people feel worse even if their returns are actually better, because on any given day there are more losing positions.

And yet, loss aversion does exist in many contexts. In business software, for example, people don't like to switch vendors for slightly-better; they will accept slightly-worse for free (e.g. agreeing to use Teams instead of Slack and Zoom). In the other direction, angel investors can't afford to think much about risk because it's just a way to refine an existing estimate that 95%+ of angel investments don't work out; figuring out whether this one's 94.2% or 95.8% is not a good use of time compared to figuring out whether the ~5% outcome is a solid strategic acquisition in a few years or a big IPO in a decade.

This creates some room for alpha creation: take a complicated economic process and you can split it into parts that are good for risk-seekers and good for risk-avoiders. And if you happen to be risk-indifferent, or can at least put on a convincing impersonation thereof, you have something to sell to both sides. Right now, bankers are busy doing this with AI. You can split an AI supply chain into different categories based on how it's financed:

1. GPUs are an asset that has some value as collateral, and there's enough of a backlog that a lender can feel reasonably confident that they'll have liquidation value in the future.

2. Meanwhile, these GPUs spit out cash flows from both training and inference. These aren't quite as predictable, but they also represent cash coming in, so a short-term loan can get rapidly paid down as a GPU's depreciation converts into inference revenue (this works even better as part of a taxable portfolio that includes more stable assets, because GPUs depreciate fast, so if they're producing less revenue than expected they're at least generating a tax loss).

3. The upside from customer-facing products is less certain, so that's a natural thing for equity to fund.

4. Even there, if a customer-facing AI company gets a decent handle on what its customer lifetime value is, there's room for a bit more credit to finance customer acquisition.

This is a big part of the function of finance. The set of cash flows produced by real-world economic actors doesn't necessarily correspond to the implicit risk preferences of different investors, and sometimes investors want to bet on a theme but don't have the right asset class to do it. With the right deal structure, risk-seekers and risk-avoiders can both get what they're looking for.

Read More in The Diff

In The Diff, we’ve talked a few times about people’s attitudes towards risk, and how financial structures can be built around this. 

• Industries naturally cycle between being asset-light and asset-heavy ($), and their financing needs change accordingly.

• One finance puzzle: why do companies have a real estate arm in the form of owning their corporate headquarters?

• One way to deal with the wrong-return-profile-for-your-preferences problem: hedging!

• The CLO boom is another instance of this phenomenon.

And note that, for a limited time, new Diff paid subscribers who choose an annual plan can get a free copy of Boom upon request.

Join the discussion!