The death and life of prediction markets at Google

&gt; Google pioneered many now standard tech practices: on-site cafés, A&#x2F;B tests, and “dogfooding,” or first releasing new products internally where they can be improved before launching to the public.<p>Famously, Microsoft and others pioneered dogfooding decades before the events described in this article and approximately a decade (at least) before Google came into existence.<p>And I’m 99% certain company cafes existed at least a half century before Google invented the concept.

I was hoping the article would reflect on the problems with predictions markets, but it&#x27;s just a dry history.<p>Crucially, &quot;predictions markets&quot; do not and cannot exist in any real sense. A pure predictions market would be completely isolated, causality-wise, from the event they are trying to predict. But the two are not and cannot be isolated, except for some degenerate cases like trying to guess the output of a true random number generator (and even then I&#x27;m not so sure sufficiently motivated people wouldn&#x27;t try to game the system anyway). This is why we have problems with our current predictions markets, e.g. the stock market (insider trading, etc.) and sports betting (match-fixing, etc.).<p>Every prediction with a stake is an incentive to alter the outcome of an event. Once the weight of the stake outweighs the resources being used to ensure the impartiality of the outcome, the wheels fully come off the cart and the prediction stops being about the underlying event and starts self-referentially predicting the impact of the prediction itself. The snake eats its own tail and the market becomes useless. You cannot scale up a predictions market without this eventually coming to pass. See also the famous example of how a predictions market for when public figures will die is just an assassination market with extra steps.

Google’s success wasn’t driven solely by its whimsical culture; the rapid growth of the market played a significant role as well!

&gt; Some questions aimed to predict its competitors&#x27; next moves, such as “Will Apple launch a computer based on Intel&#x27;s Power PC chip?”<p>Is there a point spread on journalistic accuracy? How do I take the under?

the market price gives you a number between 0 and 1, which should move higher as the event becomes more likely, so it&#x27;s pretty useful.<p>however interpreting it as a probability, or an average of agent beliefs, or anything like that, seems tricky. i assume these internal markets are not deep and liquid enough that you can just throw up your hands and say &quot;EMH&quot;. it works if you assume risk-neutral traders who will just trade up to their correct price but as I understand it, breaks down with realistic traders who may limited capital and are usually somewhat risk-averse.<p>i wonder how these prediction markets dealt with that. was there any postprocessing of market prices to get final probabilities? based on interviews with traders or observed trading behavior, did the traders behave in such a way that the market price could be interpreted as &quot;pretty much&quot; just a probability?

&gt; Yet what executives would have wanted to know was “Will Microsoft integrate LLMs into Outlook by Spring 2023?”<p>I&#x27;ve said this many times: innovation at google is dead. This is yet another supporting data point. Firms that innovate aren&#x27;t concerned with what competitors are doing. At google it appears it&#x27;s their main concern.

Great writeup. I think prediction markets will break through when they figure out how to capitalize not only on collective judgment, but super-forecasting.<p>As currently formulated, prediction market outputs are just a fancy opinion poll, where participants have some incentive for accuracy. To rise above the simple wisdom of the crowds, you would want to identify the subset of market participants that are constantly beating the market (because they have a more accurate mental model of the world). I think this necessitates both 1) long term tracking of bets and 2) likely withholding individual positions from the market to prevent follower effects.<p>Similar to the title article, this raises the question of who the ultimate customer is prediction market is. Individuals can be incentivized to bet by winnings, but who else is the customer for aggregated data?<p>I wonder about the extent to which current prediction markets have internal outputs and derivative statistics, and what they might do with it.<p>If polymarket or similar companies put Trump vs Harris at 55-45, do they have internal statistics that that put the race at 80-20% among their most accurate betters? Was this data for sale?

Google, can’t live with them, can’t do much online without them

Google hosted multiple internal gambling platforms for employees, and HR, compliance, etc. were <i>okay</i> with that?

Polymarket is the most successful prediction market imo