Adam Szymański

Problem 01 · Subsidy allocation

Every market is on life support.

Every prediction market alive today is paid to be alive. Nobody has worked out the cheapest way to keep one breathing, or where the money should go when you only have a little of it.

Mechanism design Empirical Manifold-scale budget

Simple assumes nothing

A bet needs someone on the other side. On a small market nobody is there, so the price sits still and tells you nothing. Platforms fix this by paying people to show up. Metaculus pays in prizes. Polymarket paid in tokens and airdrops. That money is a subsidy, and it comes from investors rather than from customers.

The question everyone asks is how to get rid of the subsidy. There is a smaller question nobody has answered, which is how to spend it well. If you have ten thousand dollars and want the best possible market, where does the money go? Into rewards for the people quoting prices, into a bigger prize for being right, or into paying a careful person to write a resolution rule that cannot be argued with?

Nobody knows. Every platform guesses, and the guesses have never been compared.

Moderate assumes you know what a market is

The standard framing is liquidity provision. You pay market makers to quote tight spreads, they absorb inventory risk, traders get filled, prices move on information. LMSR does this with a fixed worst-case loss, so the subsidy is literally a parameter you set before opening.

Liquidity may not be the binding constraint at small budgets. A market with a vague resolution criterion carries a wide spread for a reason no amount of maker reward touches. Traders are not pricing the event, they are pricing the chance the rule gets read differently than they expect. That is spec risk, and maker subsidy does nothing to it. My guess is that below some budget threshold, a dollar spent on writing the rule buys more price quality than a dollar spent on liquidity. The threshold should be measurable.

So this is a real optimization. Given a budget, split it across maker rewards, trader incentives, resolution quality, and how many markets you open at all, to maximize information gained per dollar. Nobody has written that objective down. There is no benchmark, no baseline, and no measured comparison of two subsidy schemes on the same event.

Technical state of the art and the gap

LMSR bounds the market maker's worst-case loss at b·ln(n) for liquidity parameter b over n outcomes, which makes b the price of a market. What is missing is any result relating b to price quality. We know exactly what a market costs. We do not know what we bought.

Three pieces are open. First the objective: a candidate is expected reduction in KL divergence between the market's terminal distribution and the realised outcome, per dollar spent, though anything defensible would be progress. Second the allocation, treating maker reward, resolution-spec investment, and market count as competing line items and finding the interior optimum. Third the exit condition. Is there a liquidity level above which organic flow sustains itself, making the subsidy a startup cost rather than an operating cost? That is a question about whether the system has an attracting fixed point, and it decides whether "self-sustaining" is even the right target or whether permanent subsidy is simply the cost of running public infrastructure.

The experimental shape already exists. Rasooly and Rozzi ran 817 Manifold markets to study whether manipulation persists. The same design works here, and Manifold will let you do it without asking anyone.

Where I would start

  1. Write the objective function down before running anything. If you cannot state what you are maximizing, the experiment produces numbers you cannot interpret.
  2. Pick one event class with recurring instances. Weekly economic releases, sports, anything that repeats. Recurrence is what lets you compare two schemes without confounding them with the event.
  3. Run the same event twice on Manifold. Once with the budget in maker subsidy, once with most of it spent on an airtight resolution spec and a smaller subsidy. Compare terminal Brier scores.
  4. Go looking for anything in the literature connecting b to accuracy. I have not found it. If it genuinely does not exist, that absence is the paper.

What counts as a result

A curve. Dollars on one axis, some measure of price quality on the other, for two or more allocation strategies. Even a rough one over a handful of markets would be the first empirical thing anyone has published here.

Related