Manifold Subsidy Experiment

A while back, Manifold handed out M$50,000 to each of a few different users, to distribute around the community in prosocial ways. (Mostly market subsidization, but a few people also used it for other stuff like rewarding community work and covering for a Manifold bug that caused people to lose mana.)

Originally, my plan was to distribute the mana via decision markets. Other market creators would pick some prosocial outcome that could occur if I subsidized their market, such as it getting lots of traders, getting it widely shared on social media, being cited by a well-known source, etc. Then they'd make markets on the likelihood of this positive outcome conditional on a subsidy, and I could choose to subsidize the ones with the largest correlation.

Unfortunately, there was little interest in this. I overestimated how much creators would care about a subsidy; I thought they'd want it to make their market more popular, since it gets them more in trader bonuses and improves their leaderboard position, but apparently that was a very minor consideration for most creators. Additionally, traders were uninterested in betting in the conditional markets, so their probabilities were uninformative.

So, I gave up on that plan and made the decisions myself. I focused on markets that were neglected and impactful; that is, ones with very few traders at the time, but were of broadly relevant interest and I thought the world would benefit from having a probability on. Some examples:

Note how all of these markets were stagnant for several weeks/months, and then as soon as the subsidy was added, they rapidly gained more than 100 traders, often bringing a change in the probability along with it.

AI is one of the most important developments going on right now, so I spent a lot of the budget on markets that aimed to get a meaningful probability on specific sub-questions about AI risk and capabilities development:

(Finding ways to put prediction markets to use on subjects with a strong utility-of-mana bias in one direction (like "everyone dies") is its own topic, and I plan to write a separate article on that soon.)

But that was only half of what I did with the budget. The other half I put towards figuring out what effect subsidies actually have on a market; do they make more people want to participate? Do they increase the incentive to predict accurately? If I want a reliable probability for some event, does it make sense for me to subsidize the market about it rather than some other approach?

I did some preliminary investigation into this in 2022I think? I can't seem to find it again. It's possible I'm confusing this with the advertisement test I did., but the results weren't conclusive. Marketwise also did their own experiment and posted this writeup, which was quite nice, but addressed a different question. (Subsidies vs. boosts.) So I decided to focus on the question of how many new traders are attracted by a subsidy of a given size.

I did this by first finding a bunch of markets with interesting enough titles that I might want to subsidize them. (Often I would find a user who had created one interesting market, and I'd look through all their other markets to find other interesting ones.) Then I'd look at the market description and decide whether it was clear enough for me to want to subsidize, and I tracked the market regardless of the decision. This meant I had some "control" markets that were superficially similar to the markets I did subsidize, but differed in some details.

I didn't like this level of confounding, so I also found some groups of almost identical markets that differed in only one variable, such as "who will win the election in [state]" or "will any country legalize recreational use of opioids before [date]". I then chose which markets within that group to subsidize entirely at random.

Numbers time! I tracked the number of traders of each market I subsidized over 2 months, to see how the subsidy affected it. I used subsidies to market liquidity only, no boosts. Some summary statistics:

(All subsidy numbers are the amount I paid into the market, not subtracting out the 25% fee.)

Restricting it to only the very similar groups of markets doesn't substantively change much. Each group was too small to do much analysis within, but the same general trends appear. Take a look at this spreadsheet for a quick summary.

My interpretation of this data is that a market's popularity is highly dependent on Manifold's market surfacing algorithm. If a market gets enough activity, it goes "viral" and starts getting shown on the homepage, which gets more people to bet on it in a compounding cycle. The larger a subsidy given to the market, the higher a chance it has of going viral.

The subject of the market is relevant; it has to be something that traders find interesting. And there's also a ton of random variance. I put M$2000 into whether a Democrat will win the state attorney general election in Pennsylvania, and got a grand total of 2 traders. While I only put M$200 into whether Qatar bring a fifth generation fighter into service, and got 50 traders. ¯\_(ツ)_/¯

If there were a way to run this experiment again and control for the popularity/activity contribution to its appearance in the market feed, I would expect subsidies to have almost no effect. I don't think many people are choosing to bet based on the available liquidity, and I don't think that available liquidity alone factors heavily into feed visibility.

This model can't explain how they get their original boost though. Subsidizing a market with >M$200 definitely leads to a sharp short-term increase in probability of going viral, and I'm not sure what explains that. Maybe the feed algorithm includes a term for "recently subsidized", and then if it doesn't get popular within a few days, it stops showing up?

My takeaway as a market creator is that if I want one of my markets to get popular, a pure subsidy is helpful, but not enough on its own. I should also put some mana towards boosting it, along with doing some manual advertising in other channels. And perhaps most importantly, all three of those things need to happen within a span of just a few days, in order to have enough concurrent activity to make it onto the homepage.

RSS feed