Design Patterns in Googles Prediction Market on Google Cloud | Hacker News Reader

Design Patterns in Googles Prediction Market on Google Cloud

cloud.google.com

64 pointsaleyan7 comments

See also Scott Alexander's comments on this, https://astralcodexten.substack.com/p/mantic-monday-let-me-g...

Edit: it's a separate HN post, https://news.ycombinator.com/item?id=29642210

Interesting, I built an app that does something similar with respect to converting forecasts to trades, allowing forecasters to trade their predictions (in hidden market) for points (not real money).

https://www.unitarity.com/app/

How does the scoring actually work? The explanation in FAQ was a bit too high level for me to make sense of it.

The problem with prediction markets is that information is not equally distributed amongst participants, and vote "weight" needs to be adjusted based upon proximity to information as well as rational, objective decision-making criteria.

Otherwise you can run into self-fulfilling prophesies, where fascination with gloom-and-doom scenarios become rampant. Imagine the "QAnon-ization" of a "predictive market." Or, conversely, the positive feedback loop that would result in a tulipmania, or the kind of crypto-coin goldbugs' "Going to the moon."

In other words, you have to account for bad faith actors in a predictive market, manipulations to it, and inherent biases.

"Incentivizing" the "right" behavior doesn't work if people are unswayable from preset biases and presumptions.

Recall the debacle with Policy Analysis Market (PAM) from 2003:

• News: https://www.nbcnews.com/id/wbna3072985

• Paper on the PAM: https://watermark.silverchair.com/itgg.2007.2.3.73.pdf?token...

(EDIT) Last note: Predictive markets tend to fail worst when trying to model for stochastic events — like terrorist attacks — which are inherently probabilistic, but not predictable. I use the analogy: "Monday, Tuesday, Wednesday, BANANA!" How would you have predicated that based on prior data?

> The problem with prediction markets is that information is not equally distributed amongst participants, and vote "weight" needs to be adjusted based upon proximity to information as well as rational, objective decision-making criteria.

This adds unnecessary complexity. Prefer simplicity. The votes are already weighted by bet size, and those with greater proximity to relevant information or insight will tend to bet more. Empirically, prediction markets have done well.

Not necessarily. I might not want to bet on my own situation. At all. Ever. Because that might tip my hand as to what I know about what's going on. Many close-to-the-problem folks might then bet "zero." But just adjacent to it, folks closer to it than a random stranger, but not as tight-lipped, might make bets.

But that's better than a "lol this rocks" or "lol this sucks" bet by someone who has absolutely no knowledge of the situation — at all. That's either going to be a random bet or, more likely, a pre-biased placing of a bet.

Also, the link I provided above expired. Here's a better link; you can then download the PDF from there:

https://direct.mit.edu/itgg/article/2/3/73/9498/The-Policy-A...

Like the article mentions, Gleangen handles the "proximity" issue by letting a predictor choose their confidence (which maps to the volume of the underlying trade). As long as people are self-aware enough to judge how confident/unconfident they are about their predictions I think this tends to work itself out in the end.

WRT to QAnonization, I managed to make decent returns on prediction markets for the 2020 election after it was over. It's like the old saying about how the lottery is a tax on people who aren't good at statistics. Prediction markets are a tax on people who have irrational beliefs and/or incomplete information.

Reply on news.ycombinator.com