aka Why (some) market forecasts are biased and what can we do about them?

tldr; When you care about an outcome, it will skew the odds you'll bet at. (To insure against bad outcomes). When everyone cares about an outcome, it skews the market odds.

People (including myself) often point at market forecasts as being the best predictors for some variable (inflation, the economy, whether England will win the Euros). 

However, market prices are more than just expectations. In the words of an anonymous twitter account:

Market price = Expectations + Risk Premium + Idiosyncratic Issues

This is a somewhat tautological framing however (since we don't know what risk premium or the idiosyncratic issues are other than the things which make this equality true). I'm going to try and explain what risk premium is, since people often focus too much on expectations and use idiosyncratic issues an excuse (fees are too high on PredictIt, you can't trade large size, etc..). Risk premium is important because unlike idiosyncratic issues which can be fixed with good market structure, risk premium will not go away. 

This matters, because if we can't remove risk premiums and the idiosyncratic issues, then prediction markets are less accurate and less valuable for decision making.

To this end, I'm going to start with some definitions:

What are risk premiums? Risk premium is excess returns which results from the payout being correlated with the rest of the world.

What are excess returns? Excess returns are returns higher than 0 (or the risk-free rate depending on context) after accounting for the riskiness of the payout.

Some examples:

  • a 1-1 bet on a 50/50 coin flip doesn't have excess returns
  • a 2-1 bet on a 50/50 coin flip has excess returns
  • a 10% 1y bond on a risky company may or may not have excess returns, depending on the default rate of the company and the recovery rate in event of default
  • the stock market has excess returns (probably)

What is correlation with the rest of the world?

When an event happens, that says something about the world state we're in as well as the payout we get from the market. If the world state affects our utility in other ways than just the market, that is the correlation we're talking about.

Some examples:

  • If I bet on a coin flip, aside from the money I have bet on the event, it has no effect on my well being.
  • If I bet on England to win the Euros, then I have some utility associated with the outcome unrelated to my bet. I will be happier if England win regardless of whether or not I bet.
  • If I insure my car against crashes then I have some utility associated with the outcome unrelated to my bet. I will be very sad if I crash my car.
  • If the economy starts doing badly, financial assets will start suffering, I will be suffering and so will everyone else.

When does correlation lead to a risk premium?

(This is me writing out with examples what Cochrane explains more clearly here in Section 1.4. Particularly "Risk Corrections" and "Idiosyncratic Risk")

What matters is the size of the market, the ability to diversify and the participants involved.

Let's start with the coin flip. If I am the only person in the world making a market in a certain coin flip, I might make the market .48 / .52 for $1 on heads. (That is if you want to buy heads, you need to pay 52c and I will buy heads from you at 48c). [This is because 0.5 * U(wealth - .48) + 0.5 * U(wealth + .52) > U(wealth) (for my current wealth and utility)] However, someone with much more money than me would look at my market and be willing to do the same trade for .49 / .51 because they are still going to earn some return. Someone with much more money than them will look at that market and offer .495/ .505 and eventually we end up with a very tight market around .50 the tightness effectively depending on how wealthy (and risk averse) the participants are.

Let's now continue with the England example. Let's say we both agree the chances of England winning are 50%, my current worth is $10, I value England winning at $10 and England losing as -$8 and I have log utility.

My current utility is 
If we bet at 55/45 (me getting the raw end of the deal) then my expected utility is 0.5 * . Therefore the market (between us) could end up clearing at the (incorrect) price of 55%. (Actually in this extreme example, I would be happy betting at up to 91%(!)). If everyone felt like me, the market might clear far from the probability everyone agrees on.

However this market has many participants and they don't all have the same (extreme!) preferences. Therefore any difference between the clearing price here and the "fair" clearing price slowly gets arbitraged away until we have a good prediction market.

What about the market for car insurance? Well, it's similar to the coin flip. I have some risk aversion associated with the cost of having to replace the cost of a car. On top of that I have the additional change in my utility which comes from having been in a car accident. So I am willing to pay more than the my estimated probability. However, there is someone else on the other side of this. When I go to renew my insurance there are many companies offering me car insurance. They are in competition, each with a large float and indifferent to whether or not I crash my car. Their only interest is their expected return. This competition drives the price closer to the true probability. (More on this later...)

What about equities? Well it's similar to the coin flip. Uncertain money is worth less than certain money to me. It's also similar to the insurance example. More money in a good economy is less good than less money in a bad economy is bad for me. However, here we need to start thinking about the other side. Who is the other side of the equity market? Well, there are certainly people with much deeper pockets than me, so "uncertain money" doesn't drive them as much so the uncertainty value gets priced out. However, there are very few people out there who have pro-cyclical utility - (people who want more money in good times than they want less money in bad times). So the market clearing price (as a whole) will be less than the "fair" price. (This is one "solution" to the Equity Premium Puzzle).

What does this mean for prediction markets?

Well for simple prediction markets: coin flip; sports events; weather(?); there are sufficiently many people who are either indifferent enough, or deep-pocketted enough that they will take on the risk for an ever decreasing edge. This means the market price should tend to the market participants' best estimate of the fair probability.

For complicated markets: insurance, equities, inflation-protected bonds, it's not quite so simple to infer a probability/forecast/expectations from market prices.

Let's start with insurance. We know that the probabilities which we infer from the insurance market are (broadly) an overestimate. (Insurance companies are profitable after all!) This would be fine if we knew how much of an overestimate they were. "Take the market prices... subtract off the profit margin from the insurers... done". Unfortunately, the insurance market is slightly more dynamic than that. (See the time varying loss ratios from this fascinating set of slides

Over time the amount of risk premium being demanded by the insurance companies waxes and wanes. Therefore we need to know the right adjustment to make for risk premium. (And arguably this is just as hard as predicting prices in the first place - no free lunch here).

What about equities? If we know that equities broadly go up, can't we just subtract off that positive rate of return to use them as indicators for the real economy? Again, no. The problem is that the risk premium is time varying and knowing what the current risk premium is, is about as hard as valuing equities in the first place. No free lunch here.

What about TIPS? TIPS are the classic example of using the financial markets as a forecasting tool. They are inflation-protected bonds issued by the US government. In theory, the difference between TIPS and nominal bonds will be the market's expectation for inflation. Certainly the difference* will realise to inflation. The question is, will the market clear at that price? The empirical evidence seems to be "not quite" **.

What about GDP-linked bonds? Well, I can't think of anything more correlated to the real economy, so I would expect the risk premium of these (for global economies) to be very large (relative to their volatility). This would compare to nominal bonds which have a negative risk premium. This means for the country issuing them, they would be much more expensive than traditional debt issuance.

Their use as a forecasting tool would also be weakened, since they would systematically under-forecast GDP growth.

That said - I am still a fan of the idea. I would rather implement them as futures contracts (pretty much what Sumner advocates for here).   

How bad can it get?

We can take some hope (or despair) from the Hansen-Jagannathan bound. This is usually used to say something about how volatile the state price density**** is. However, this volatility gives us an upper-bound for the Sharpe of any asset. (Including any asset we might use for prediction). If we are willing to assume*** that empirically the upper bound for the state price density is 0.5 then we know that . So there is at least hope that market prices will get us somewhere within an order of magnitude of the right answer.

* modulo some details regarding options

** there are a bunch of different factors driving this difference, some premium related some market inefficiency related, some to do with large flows in the market

*** this is a key assumption, but I'm quite willing to believe it, since no-one has found assets with much higher Sharpe than 0.5

**** For more information on the state price density and why this is relevant, the chapter of Cochrane linked at the start should be a good first port of call.

New Comment
6 comments, sorted by Click to highlight new comments since: Today at 4:27 AM

Idea for dealing with the risk premium hedging problem:

A prediction market that requires participants to make two predictions on each question, one in a real money market, and one where they just get points like metaculus. 

Then if there's no difference, I'm pretty sure you can be confidant risk premiums and hedging didn't have an effect. If there is a difference then it's a bit more complex to know why that is (It's not necessarily because of hedging, could be that people just don't care about the points), but it can give a hint that there's hedging, especially if the difference comes from players who have a good track record in both categories. If you think you know why the difference is there then you can use the results from both markets to get better information for decision making.

Do you think something like this could work?

I don't really know how you incentivise people (seriously) in the non-real money prediction markets.

Non-money prediction markets have lots of difficulties to them:

  • How do you size your bet? (Ie knowing a probability vs "higher" or "lower" than market estimate
  • Difficult to arbitrage (ie share information between markets)
  • How do you show your conviction (this is 50% and I'm certain it's a coin flip vs this is 50% because I don't understand the question)

Assuming Sharpe tops out at 0.5 seems wrong. Over the last 5 years, Bitcoin has had a Sharpe ratio of 1.6 the Wilshire 5000 has had a Sharpe of 1.04.

Your second example (to me) demonstrates the issue with your first. If over a 5y window you can be convinced that US equities have a Sharpe >1, you can be convinced over a 5y window that BTC has a Sharpe > 1.5. 

[The Sharpe of the best performing stock is always going to be phenomenally high, does that refute this - I say no, on the grounds that we're interested in ex-ante Sharpe. If the strategy "always buy the best performing stock" had a Sharpe ratio of >> 0.5 then I would be interested, but I don't believe that is what we see.]

To be clear, I don't believe that Sharpe (in general) tops out at 0.5. I think there are plenty of strategies which have Sharpes >> 0.5. I just don't think they are especially relevant in this context. (Usually because they are capacity constrained, although there are other reasons).

To be honest, if you don't find 0.5 convincing - that's fine. You should have even less faith in the value of prediction markets. Nothing wrong with that - it's the whole point of my post. I just wanted to give people a little hope (or at least the hope I have).

I was also surprised to read that because I was always under the impression that the SR of bonds is higher than that of stocks over a long time, and the SR of the stock market over a long time (much longer than 5 years) is about 0.5. 

However, when I looked into the data more closely, the recent (1976-present) high sharpe ratio of bonds seems to be an extremely temporary phenomenon related to the large decline in interest rates during this time. The SR of bonds prior to 1976 is actually worse than stock index SR. So when looking at a very long timespan (on the order of about 100+ years), I actually agree with SimonM on this upper bound for assets now. 

At the very least, I do not know of any asset classes with 100+ years of history that have SR over that period significantly greater than 0.5

One of my favourite papers The Rate of Return on Everything suggests that property might have Sharpe ratios > 0.5 with 100+ years of history, although I generally think that the volatility is somewhat underestimated there. (It also has a lot of data on the Sharpe ratios for long bonds + equities across a wide range of DM for long histories)