robin hanson's logarithmic market scoring rule is the one piece of prediction-market machinery that almost everyone agrees works. it solves the bootstrap problem — a single market maker can quote a price before any order arrives, and the worst-case loss is bounded by a small parameter b that you choose. for a single well-specified question with a curious audience, lmsr is the right answer and has been for almost twenty years.
but if you have spent any time actually trying to run a prediction market, you know the problem is not one market. the problem is a catalog of thousands of them, almost all of which will attract single-digit numbers of traders, and whose combined subsidy bill grows with the catalog. lmsr is designed for the shape of a market. the catalog is a different mechanism-design problem, and the people building prediction markets have mostly not noticed this.
here is the issue in one paragraph. lmsr's worst-case loss is b · ln(n), where n is the number of outcomes. that is fine for one market. scale to a thousand markets and your subsidy budget is a thousand times b · ln(n). the return on that subsidy is the information value of the aggregated prices, which is heavily concentrated in a handful of high-traffic markets (elections, big sports events, fed decisions). the long tail — "will company x miss its q3 numbers," "will bill y pass by december" — generates roughly zero information value per subsidy dollar because nobody trades them. you are paying a market-maker cost for a market that is functionally a survey of two people.
the usual response is "so lower b." but if you lower b, spreads widen and slippage on any non-trivial trade becomes prohibitive. you have moved your failure mode from "subsidy too expensive" to "prices too noisy to inform anyone." there is no b that solves both problems at once for a long-tail market, because the long-tail market genuinely does not have enough volume to support a narrow quote without burning the market-maker.
the fix, i think, is to stop treating the automated market maker as the pricing mechanism and start treating it as a fallback. the pricing mechanism should be a batched auction: collect orders over some interval (minutes, not milliseconds), clear at the price that maximizes matched volume, and fall back to the amm only when there is no two-sided interest. batched auctions kill hft-style adverse selection, which is most of what makes thin books expensive to quote, and they let the catalog subsidize many markets with one pot of liquidity rather than n pots.
a two-tier design gets cleaner when you think of it as a contracts market sitting on top of a shared liquidity reserve. the reserve provides the lmsr-style guarantee that a trader can always get a quote. the per-market batch auction provides price discovery when there is enough interest to clear natively, which avoids paying the market-maker cost on every trade. thin markets default to the reserve and absorb the expected subsidy loss pro rata. thick markets rarely touch the reserve because their own order flow is enough.
there is a second, subtler benefit to batched clearing for long-tail questions: it forgives the absence of continuous attention. a retail trader who thinks "will x happen by december" is 20% likely does not want to babysit a limit order for weeks. a batch auction with a settlement every hour or every day aligns with the cadence of casual forecasting. lmsr-style continuous quoting implicitly assumes traders can keep up with the book, which is true for sports and elections and false for everything else.
what this looks like in practice is boring. you run a standard clob per market with batched rather than continuous matching. you let an amm sit in the background as a liquidity of last resort. you subsidize the amm's worst-case loss at the portfolio level — one budget for the whole catalog — not per-market. and you admit, up front, that most markets in the catalog will never clear interesting volume and are being kept around for the option value of letting anyone ask any question.
i suspect the next few years of prediction-market design will be dominated by this split. polymarket has the catalog but uses a single amm per market, and the long-tail problem is visible in their book depth. kalshi is more centralized but has spent enormous engineering effort on matching, which is suggestive of where they think the bottleneck is. metaculus does not have money at all and is mostly solving the attention problem rather than the pricing one. none of them, to my knowledge, has put real work into the catalog as the object of design rather than the market.
if you are building this, the summary is: lmsr is a brilliant single-market primitive, not a platform primitive. the platform needs batched auctions, portfolio-level subsidy budgets, and a liquidity reserve whose job is to make rare markets quotable rather than to price them. everyone is still running lmsr end-to-end and wondering why the long tail is empty.