There are a number of features that Manifold could implement that would improve the accuracy of markets at relatively extreme probabilities (approximately, > 95% certainty). This bounty is a place to collect my notes on them. Bounty awards start at M500 for relevant suggestions, or M5000 for accepted pull requests that implement them. If you'd prefer payment in some non-Mana currency, please let me know in advance. Proposals and requests for larger bounty awards are welcome. (This bounty is underfunded; I'll add more as needed. Contributions welcome.)
See also:
https://manifund.org/projects/extreme-probabilities-project
I believe there are multiple related causes, all of them tractable if difficult to fix in their entirety. First and foremost, this is an interest rate and capital costs problem. Manifold's loan program is an excellent start in this direction, as it allows users betting on the "expected" side of a lopsided market to not tie up their capital for the entire duration of the market.
Further increases in the loan rate beyond 4%/day would help more. However, they come with increasing risks to Manifold from holders of large loans potentially defaulting, or large numbers of small users (likely new ones, attracted by a popular market) becoming over-leveraged on individual markets. So, a feature request: higher loan rates for well-diversified users, or other methods of risk management on faster loans.
Similarly, many of the markets of interest take the form of "will an event happen by a time", and the expected market behavior is for the likelihood to decay over time as the event continues to not happen. This means that people betting the event won't happen are making a profit, but they do not get that profit loaned back as they would with their initial investment. Paying out loans on unrealized gains also has risks, especially if there is no symmetric requirement to repay unrealized losses.
The interest rate problem also shows up in places where capital requirements exceed what they need to be, in that a user is required to tie up more capital than they can possibly lose. For example, when holding multiple No positions on a multiple choice market, or when the markets are logically dependent.
There are also a few minor obvious ones, such as being able to bet more precisely in extreme cases. Restricting probabilities to integer % values in ordinary ranges is bad, but low impact; at extreme ranges there is a meaningful difference between (say) 2.5% and 2%, or 2% and 1%. At least by the time you get to < 3% or > 97%, finer gradations of probability should be available, including values more extreme than 1%. There is a reasonable case that these should often be displayed as "< 1%" or similar in order to avoid presenting misleading claims of precision, but the actual value should be available and it should be possible to set limit orders at these values (especially important for arb bots).
I'll continue adding to this list over time.
Related questions
Scenario, new feature introduced:
Questions A, B, C, D, E, F, G, H, I, and J are each 1% YES questions.
Someone buys 100 shares of NO on each of them, for a total of 990M. However, rather than having all that mana tied up, the user only puts in, say, 250M as collateral for the combined 1000 shares.
If A resolves YES, then 99M is lost from the collateral (bringing it down to 151M), and then shares each of the remaining questions are automatically sold until the M/value ratio is back: ~8 shares of each, so the investment is left with ~205M and ~92 shares each of the 9 remaining questions.
And so on for the remaining questions if they resolve YES (or have the value of the NO shares drop sufficiently), perhaps also keeping the mana left above a minimum of a certain multiple of the amount of risk in the set, such that it can't ever drop below 0 in total value.
There's probably a word/system for something like this in IRL trading, but I do not know what it is.
Not sure if this is a good suggestion, but since more traders usually make a probability more accurate, the number could reflect that. At extreme probabilities, it could show 0 decimal points if it has less than 20 traders for example, and then have 1 decimal point once it reaches 20, maybe even 2 decimals when it reaches 100, something like that.
Manifold could add a built-in way to make "amplified" markets like this one. This would be helpful for making precise bets on markets that you expect to have a very low probability.
Hmm, what about having a second pool of cash "extremebucks" or something, that can only be invested on questions with >95% or <5% odds, and this is the only way to convert them to regular mana? Might be clunky, but it would force people to bet on these extrema. I don't know how you would stop people making questions like "is 2+2=4" to convert them though.