Being Right Is Overrated

**by **@allquantor

The Error Maximization Machine

People in finance like to think they are in the business of predicting the future. If you run a hedge fund, or trade crypto, or bet on prediction markets, your self-conception is probably that you are a person who knows things. You have a model, or a feeling, or a guy on the inside, and that tells you that Apple is going up or that the yield on the 10-year Treasury is going down. You are in the business of being right.

But there is a different view of finance, one that is less about being a wizard and more about being an actuary. This view says that knowing the future is actually not that important. It is arguably optional. What matters is how you structure the bets.

There is a huge pile of academic research on this, ranging from boring studies of pension funds in the 1980s to exciting studies of coin-flipping monkeys in the 2010s. The conclusion of all of it is roughly this: You can be a genius predictor and go broke, and you can be a mediocre predictor and get rich. The difference isn't the prediction. It is the portfolio management.

The 90% Rule

In 1986, Gary Brinson, Randolph Hood, and Gilbert Beebower published a paper that really annoyed a lot of stock pickers. They looked at 91 large pension plans over a decade. These plans spent millions of dollars on "active management." They hired very smart people to decide which stocks to buy (security selection) and when to buy them (market timing).

The researchers ran the numbers and found something embarrassing. About 93.6% of the variation in the plans' returns was explained purely by their investment policy.

That means if a fund decided to be "60% stocks and 40% bonds," that decision basically dictated their entire financial life. The active decisions picking Apple over IBM, or selling before a crash accounted for almost nothing. Actually it was worse than nothing. On average, the active management cost them money.

This is the "90% Rule." It says that 90% of your returns come from just showing up in the right asset classes, and the other 10% comes from your frantic attempts to be clever.

Later researchers like Roger Ibbotson refined this to point out that while asset allocation explains your wealth level (whether you retire rich), active management explains the difference between you and your neighbor. If you want to get rich, you need a good allocation policy. If you want to feel smarter than your friend Bob, you need to pick stocks. Most of the financial industry is built on the desire to beat Bob, not the desire to get rich.

The Error Machine

"Okay," you say, "I won't pick stocks. I will just feed my predictions into a computer and let it optimize the portfolio."

This is also dangerous. The standard way to do this is Mean-Variance Optimization. You tell the computer your expected returns, and it gives you the mathematically perfect portfolio.

The problem is that the computer takes you seriously. If you tell the computer that Stock A will return 5% and Stock B will return 5.1%, and they have the same risk, the computer might tell you to put all your money in Stock B and short Stock A.

But you don't actually know that Stock B will return 5.1%. You are guessing. If your guess is off by 0.2%, the "optimal" portfolio is a disaster. Richard Michaud famously called this process "error maximization." The optimizer takes your tiny statistical errors and leverages them until they blow up your account.

The fix for this is "shrinkage." You basically tell the computer: "Look, I think Stock B is better, but I am probably wrong, so don't do anything crazy."

A study by Ledoit and Wolf in 2004 showed that if you just ignore some of the data and force the model to be more boring (shrinking the covariance matrix), you can increase your risk-adjusted returns by something like 50%. You don't need a better crystal ball. You just need to stop trusting the one you have.

Being Wrong for Profit

This problem is even funnier in modern machine learning. The usual way to train an AI is to minimize "prediction error" (MSE). You want the AI to guess the stock price as accurately as possible.

But a paper by Elmachtoub and Grigas points out that this is often irrelevant.

Imagine you predict a stock will go up 1%. It actually goes down 1%. Your error is small. You were close! But you bought the stock, so you lost money.

Now imagine you predict the stock will go up 100%. It actually goes up 20%. Your error is huge. You were way off! But you bought the stock, so you made money.

Academics call this "Decision-Focused Learning." It turns out that having a model with a low error rate is not the same thing as having a model that makes money. Cenesizoglu and Timmermann looked at this and found that the relationship between statistical accuracy and economic value is "weak."

You can have a "bad" model that makes money because it gets the sign right on big days, and a "good" model that loses money because it is precise about the wrong things. The alpha is not in the belief. The alpha is in the mapping from belief to action.

The Coin Flip Disaster

The best evidence that humans are structurally incapable of handling an edge comes from a 2016 experiment by Victor Haghani and Richard Dewey.

They took 61 people, including finance professionals, and gave them $25. They let them bet on a coin flip.

Here is the kicker. They told them the coin was biased. They explicitly said, "This coin will land on Heads 60% of the time."

This is the dream. You know the future. You have inside information. You should just bet on Heads, size it according to the Kelly Criterion (about 20% of your bankroll), and print money.

What happened?

1. 28% of the participants went bust. They lost all their money on a game where they had a massive, guaranteed edge.

2. Some people bet everything on Heads. They hit a streak of tails and died.

3. Some people bet on Tails. They believed in the "Gambler's Fallacy."

4. Most people just bet random amounts based on vibes.

The authors concluded that "bet sizing is not intuitive." I would conclude that if you give a trader a money-printing machine, he will find a way to stick his hand in the gears.

The study proves that prediction is useless without structure. Knowing the coin is 60/40 is the prediction. Betting 20% of your bankroll is the structure. If you have the prediction but not the structure, you go bust. If you have the structure but no prediction (e.g. betting small amounts on a fair coin), you just break even. You need the structure to survive long enough for the prediction to matter.

Survival of the Fittest

The synthesis of all this research is that Portfolio Management is not just a safety rail. It is the engine.

Richard Grinold has a "Fundamental Law of Active Management" that puts it in an equation:

Your performance (IR) equals your skill (IC) times the square root of how many bets you place (Breadth).

Most people spend their lives trying to increase their skill. They want to be smarter. But the math says it is exponentially easier to just place more, smaller independent bets.

If you are betting on prediction markets, or buying stocks, or running a pension fund, you probably spend 90% of your time thinking about "what will happen." The math says you should spend 90% of your time thinking about "how much to buy."

**The prediction is a fragile opinion. ** The allocation is the machine that allows you to survive being wrong long enough to eventually be right.