In 1994, the spread on Apple Computer common stock was twenty-five cents — not because the underlying volatility justified it, but because no NASDAQ dealer would post an odd-eighth quote. William Christie of Vanderbilt and Paul Schultz of Ohio State documented the phenomenon in The Journal of Finance that December, showing that odd-eighth quotes were "virtually nonexistent" for 70 of the 100 most actively traded NASDAQ securities, including Amgen, Cisco, Microsoft, Lotus, and Apple. The mechanical implication was a minimum inside spread of $0.25 on names that trade by the millions of shares per day. The Christie-Schultz hypothesis was that NASDAQ dealers were implicitly colluding. The day after national newspapers ran the story on May 26-27, 1994, dealers in Amgen, Cisco, and Microsoft sharply increased their use of odd-eighth quotes, and effective spreads collapsed by roughly fifty percent.
What followed was a fifteen-year compression of dealer rents. The Department of Justice and the SEC opened parallel antitrust and securities investigations. Private litigation against thirty-three of the largest NASDAQ market making firms produced a settlement north of one billion dollars. The SEC's January 20, 1997 Order Handling Rules forced market makers to display customer limit orders inside their own quoted spreads — a change that, combined with the rise of Island, Instinet, and Archipelago as electronic communications networks, exposed customer interest directly to one another rather than mediating it through a dealer. Decimalization in early 2001 finished the job. The minimum tick on a $100 stock fell from 6.25¢ to 1¢, and the legacy NYSE specialists — LaBranche, Spear Leeds & Kellogg, Fleet Meehan, Van der Moolen, Wagner Stott Bear — found themselves trying to earn a penny on quote-driven obligations that used to pay an eighth or a quarter. By 2004 the SEC had extracted more than $240M from the five biggest specialist firms for trading-ahead violations. Within the decade, the specialist post on the NYSE floor was a museum piece.
The successor industry rebuilt the franchise on different physics. Citadel Securities, Virtu Financial, Jane Street, Hudson River Trading, Susquehanna International Group, Jump Trading, and Tower Research earn the spread on millions of trades per day with millisecond-scale inventory turnover, pennies of edge per fill, and risk limits that scale with realized volatility. Knight Capital — the last of the great NASDAQ wholesalers, executing roughly 40% of online brokerage trades by 1999 — was destroyed in 45 minutes on August 1, 2012 by a software deployment bug, and the rest of the wholesaler franchise (Mayer & Schweitzer at Schwab, Herzog Heine Geduld at Merrill) was absorbed or wound down. The economic insight that survived from the 1990s is that market making is, statistically, the highest-Sharpe strategy in any electronic order book, because the maker is the counterparty to net retail demand. The economic insight that the 1990s also taught is that maker rents are an artifact of market structure — collusive eighths, opaque ECN routing, slow customer-order display — and any structural reform that increases competition or transparency compresses spreads toward the cost of capital.
The theoretical floor under any serious market-making analysis is two papers from 1985. Lawrence Glosten and Paul Milgrom's "Bid, Ask and Transaction Prices in a Specialist Market with Heterogeneously Informed Traders" (JFE, 1985) showed that even a risk-neutral, zero-profit dealer must quote a positive bid-ask spread when some fraction of order flow is informed. The mechanism is conditional expectation: when a buy order arrives, the dealer rationally revises the expected fundamental upward, because informed traders only buy when they know the asset is undervalued. The optimal ask therefore sits above the unconditional mean by exactly the expected revision conditional on a buy; the bid sits symmetrically below. The width of the spread is set by the ratio of informed to uninformed flow and the variance of the fundamental. Albert Kyle's "Continuous Auctions and Insider Trading" (Econometrica, 1985) generalized the same intuition to a strategic single informed trader who optimizes execution against a market maker observing only aggregate order flow. The equilibrium pins down "Kyle's lambda" — the slope of price impact in net order flow — as proportional to the ratio of fundamental volatility to noise-trader volume. Lambda is the empirical handle: a market with high lambda is one where every dollar of order flow moves price, which means the maker is exposed to large adverse-selection losses and must quote wider to survive.
Together, Glosten-Milgrom and Kyle imply two operating constraints. First, the spread must cover expected adverse selection — a function of how much private information traders are likely to possess. Second, the spread must cover inventory holding costs — the maker has to mark-to-market a position they did not want, and their willingness to hold it is bounded by their capital and risk limits. The equilibrium spread is the maximum of these two constraints. In equity markets after decimalization the adverse-selection component is small for liquid names because flow is mostly index funds and quantitative rebalancers; the inventory cost is small because the holding window is milliseconds. Spreads collapse to single-digit basis points. In prediction markets neither condition holds.
Polymarket runs an off-chain central limit order book settled on-chain on Polygon, with maker and taker addresses both posting USDC collateral and a smart contract paying the winning side at resolution. The Liquidity Rewards Program is documented at docs.polymarket.com/market-makers/liquidity-rewards and is essentially a port of dYdX's perpetuals incentive design with adjustments for binary markets — distinct order books per outcome, no staking, no protocol token, USDC paid directly to maker addresses. Two parameters per market govern eligibility: min_incentive_size, the smallest order the program will score, and max_incentive_spread, the widest distance from the size-cutoff-adjusted midpoint at which an order earns rewards. Both are exposed via the public CLOB API. The scoring function is quadratic in spread, which is the load-bearing detail: moving an order one cent closer to mid produces an outsized increase in score, so the rational LP quotes as tight as their adverse-selection budget allows. Two-sided quotes are weighted more heavily than one-sided, and depth is rewarded linearly subject to the spread cap. Polymarket has allocated over $5M per month in LP incentives in April 2026, concentrated in sports and esports markets where the resolution risk is mechanical (final score from a regulated league office) rather than interpretive. The "Sponsor Rewards" feature lets any user deposit a fixed USDC pot that is auto-distributed to LPs in a chosen market over a chosen window — a primitive that allows market creators to bootstrap their own books without waiting for protocol-allocated rewards.
The rewards are non-trivial in absolute dollars. Early reports from LPs concentrated in two-sided sports books described $200-300 per day on $10K of working capital — roughly a 1-3% daily yield, annualized into the hundreds of percent if it were stable. It is not stable. The same LPs note that rewards became "a thin bonus on top of real trading edge" because the adverse-selection draw on a Polymarket book is structurally larger than on an equity book of equivalent depth. A regulatory news headline, a poll release, a missed shot in a basketball game, or a Truth Social post can move a contract twenty cents in seconds. A maker quoting a one-cent spread around a $0.50 contract is exposed to a 20-cent jump every time news arrives faster than they can pull. The serious LP stack therefore looks nothing like equity HFT. It looks like a news monitoring system (Reuters, AP, Bloomberg terminal, social-media feeds, sport-specific data feeds), an automated quote-pull layer (yank all resting orders within milliseconds of any flagged headline or volatility spike), an inventory-management layer (target-flat to a per-market dollar limit, with hedges into related contracts where they exist), and a model that estimates the implied lambda for each individual market and sizes the spread accordingly. The retail "set it and forget it" reward-farming strategy is dominated by a $50K/month engineering investment for any LP working at scale.
The institutional curve is now visible. On April 3, 2024, Susquehanna International Group launched a dedicated event-contracts trading desk and announced a market-making relationship with Kalshi — the first traditional Wall Street market maker to commit to a U.S. CFTC-regulated prediction-market venue. Kalshi co-founder Tarek Mansour framed it on X as the platform finally surmounting "their most elusive challenge: liquidity." Susquehanna's edge in event contracts is the same edge that built it as one of the world's largest options market makers (~$2T of ETF volume per year): probabilistic pricing of contingent claims, with internal models that output a fair-value distribution rather than a point estimate. On February 9, 2026, Bloomberg reported that Jump Trading is set to take small equity stakes in both Kalshi and Polymarket in exchange for liquidity provision — a fixed equity grant on the Kalshi side and a stake on the Polymarket side that grows with provided trading capacity. Jump has reportedly recruited a team of more than twenty traders for the event-contracts business in recent months. Robinhood-Susquehanna's 2024 acquisition of LedgerX gave them a vertically integrated derivatives venue. The picture is consolidating along the same lines as equity wholesaling in the early 2000s: three or four firms with proper risk infrastructure, a fringe of smaller participants, and a long tail of retail LPs collecting rewards.
The most important empirical paper on prediction-market microstructure is Akey, Grégoire, Harvie & Martineau's "Who Wins and Who Loses In Prediction Markets? Evidence from Polymarket" (SSRN 6443103, March 2026; discussed on the Harvard Law School Forum on Corporate Governance). The dataset is comprehensive: more than 1.4 million users from 2022 to 2025, more than $20 billion of volume, over 70 million trades. The authors construct per-trader regressions that control for capital deployed, trade frequency, contract category, and time period, and then ask which behavioral attributes predict whether a trader is in the green at the end of the sample. The single largest predictor by economic magnitude is maker-versus-taker status. Moving from a pure taker — someone who lifts existing offers — to a pure maker — someone who only ever posts limit orders that get hit — reduces the probability of losing money by roughly 36 percentage points. The headline distributional facts are equally bracing: the top 1% of traders capture 84% of all gains, the top 0.1% capture 58.5%, 70.8% of users are net losers, and 63% of all trades execute at extreme prices below 10¢ or above 90¢ (the "lottery ticket" segment that Bürgi-Deng-Whelan and Le 2026 also identify as the structural sink for retail capital). Market making, in other words, is not just a high-Sharpe strategy — it is the strategy that flips the prediction-market base rate from a 70% loss probability to roughly even money, before any further edge is added.
The second-order economic question is why, given those returns, makers do not pile in until the edge is competed away — as happened to NYSE specialists between 1994 and 2004. The answer is the resolution risk premium. Cross-platform Kalshi-versus-Polymarket spreads on identical underlying questions persist at 1.5-4.5% even when both books are deep and liquid. The Cardi B Super Bowl halftime contract resolved YES on Polymarket and NO on Kalshi for the same factual event, because Kalshi's settlement specification required "performing" under stricter criteria than Polymarket's UMA-arbitrated definition. A U.S. government-shutdown contract divided the same way. The April 2026 Khamenei status contracts diverged because the platforms specify different resolution sources for the same underlying question. A maker who arbitrages a 3% Kalshi-Polymarket spread is taking a binary basis risk that, in the event of settlement divergence, costs the full notional. As Mansour put it on a recent podcast: "it's not like they can spin up a new desk to price politics or price culture in an hour." Equity market makers do not face this risk because the SEC and DTCC specify a single canonical settlement for every U.S. listed security. Prediction-market makers face it on every contract that resolves on a phrase rather than a number.
The practical guidance for a serious participant is therefore a hierarchy. First, the maker stack: low-latency exchange connectivity, an automated quote-pull triggered by volatility spikes and news flags, an inventory engine that targets per-market dollar limits and hedges into liquid neighbors, and a Glosten-Milgrom-style spread model that widens with realized lambda. Second, market selection: sports and esports markets reward LP because the resolution is mechanical; political and cultural markets pay wider spreads because the resolution risk is interpretive and irreducible. Third, capital scale: rewards programs pay enough to cover incremental capital costs only if a serious model is doing the inventory work; they are not a substitute for trading edge. Fourth, the institutional reality: Susquehanna and Jump are competing for the same liquid books, and the open seats for independent LPs are in mid-tier markets ($100K-$5M of volume) where the institutional desks have not yet built specialized models. The sustainable retail edge is rewards-farming on niche sports markets where two-sided depth is structurally thin and where the LP can monitor the underlying league feeds faster than the median taker. Anything more ambitious requires the same engineering and risk infrastructure that destroyed the LaBranche franchise — and earns it, on the new venue, until the next Christie-Schultz paper arrives.