Liquidity vs Promotion: Both are useful on Manifold
Ever made a prediction market and they ignored it? Manifold has two ways to buy attention and they sponsored a little experiment to compare them.
Manifold allows you to gift money to a market to encourage more trading (subsidizing). It also allows you to spend money on advertising markets (boosting). Is that worth it?
I got 50,000 Mana to play with, so I that sounded like a good opportunity to do a little research about this.
Methodology
My approach was to find 10 underrated questions and invest 5000 Mana on each. Looking for questions takes time, so ten was a good number. I doubt I would have found twenty.
An underrated question here means:
Less than 10 traders.
Resolves soon (for quite a generous meaning of “soon”).
I’m not the creator myself (although those are obviously the most underrated questions of course).
Look for important questions, often geopolitics. No sports or jokes.
Avoid US questions because those are already at an attention advantage anyways (and I care more about Europe).
Randomly decide if the question gets only a 5k subsidy, a 5k boost, or a mixed 2500 subsidy + 2500 boost.
For each question, I also picked a control group question with the same number of traders and mostly similar in other aspects. I even left a comment there, because I left a comment on the subsidized questions too.
After a week, I told Joshua about the subsidizes, so they got the Subsidy Spotlight tag and some additional publicity maybe.
Here is my Google Sheet with all the data. There are two interesting metrics, I recorded there:
How many traders participate? More diversity should correlate with more accurate predictions.
How high is the volume? High liquidity means people can earn more money, so there is a higher incentive to predict correctly.
What went wrong
In one case I forgot to actually subsidize the question, but that was easy to fix.
One question already resolved. That was too soon.
I accidentally subsidized a control group question. It is a very “scientific” one and because of that very people participate. It is quite liquid now though, so worth a look?
Tracking Over Time
Let’s look at the volume first. The “subsidy only” questions get the biggest volume boost of 5000 Mana. The “sub & boost” only gets 2500 Mana liquidity.
Surprisingly, two of the three red “subsidy only” questions stay quite low. This suggests a big subsidy without promotion is not the optimal approach.
One question which resolved within less than a week got to 3k Volume. It would have been a strong one too.
Looking at the traders, we can clearly see that the blue “boost only” did gain traders indeed. The red “subsidy only” did not. The mix was all over the place but generally strong and not too much behind the blue ones.
Here is the Google Colab notebook I used to generate the plots.
Boost AND Subsidize!
My hypothesis was that subsidizing should help a lot, but boosting not really. I heard this opinion from others too. A boost looses its effect once used up. Subsidies provide liquidity until the market closes.
Now I think that boosting does indeed help to get more traders and it does combine well with subsidies. So using both seems to be the most reliable strategy.
There is always future work, of course. Maybe splitting your investment in half is not the best approach? It could be another experiment to use a 20/80 split or 40/60 or something.
Another interesting question is the subsidy amount. M5000 is a lot considering that a question starts at M100. Looking at the plots you see that after two weeks half the questions still show less volume than subsidized.
May your markets be interesting even without boosts and subsidies!
I would appreciate some feedback. Found an error? Have more questions? Different interpretation?
Very cool! It sounds like you selected paired control group questions after selecting the main question that would receive the subsidy; presumably this means that the control group was systematically less interesting / deserving of the subsidy?
An approach you could try is to select pairs of well-matched questions, then within a pair randomize which is the control question and which gets the subsidy / boost. (You could also try things like boost one / subsidize the other, but at small N that seems like it would mostly just make the math a mess.)