In November 1990, Ken Griffin opened a Chicago shop called Wellington Financial Group with $4.6 million — most of it staked by Frank Meyer of Glenwood Capital — and pointed it at one strategy: convertible bond arbitrage. He was twenty-two. The pitch was simple to state and difficult to execute. A convertible bond is a composite security: a corporate bond glued to an embedded equity option. Through the 1980s and into the early 1990s, the desks that issued and traded these instruments — at Drexel, Salomon, First Boston, the Japanese warrant houses — priced the embedded option with what one veteran later described as "calculators and intuition." Griffin had built proprietary software that priced the option more accurately, and a real-time data infrastructure to feed it. In 1991 his book returned 43%. In 1992 it returned 40%. In 1994 the firm was renamed Citadel.
The structural insight Griffin operationalized is older than Citadel and more general than convertibles. A composite security is a basket of cash flows the market does not always know how to add up. When the basket trades cheaper than the sum of its priced components, you buy the basket and short the components. When the basket trades richer, you do the reverse. The trade is not directional; it is a bet that the market's pricing of the whole and its pricing of the parts will eventually reconcile. Convertibles in 1990 were the cleanest instance because the parts were unambiguous — a straight bond plus an equity call — and the dealers' option pricing was crude. Long the convertible, short the right amount of underlying stock, dynamically rehedge as delta shifts, and you are net long realized volatility against implied. The position has positive gamma and positive vega. It loses money in a vol crush. It bleeds carry in a quiet tape. But against a long book of cheaply written corporate optionality, the math compounds, which is what Citadel demonstrated for a decade.
The structure of multi-outcome prediction markets is a near-perfect re-instantiation of that puzzle. A Polymarket or Kalshi market with N mutually exclusive, collectively exhaustive outcomes is, by no-arbitrage, a composite asset whose constituent YES contracts must sum to exactly $1.00. Buying one share of every outcome guarantees a $1.00 payout, because exactly one resolves YES. If the sum of the asks is below $1.00, you have a complete-set arbitrage: buy the full basket for $0.995, collect $1.00 at resolution, lock in 50 basis points. If the sum of the bids is above $1.00, you have the inverse: sell every outcome short for $1.05, pay $1.00 at resolution, pocket 5 cents. These are not analogies to convertible arb. They are convertible arb's defining mechanic — pricing the parts incorrectly relative to the whole — translated to a venue where the "parts" are binary contracts and the "whole" is a tautology.
A worked example from Polymarket's Fed-decision multi-outcome family makes the arithmetic concrete. In an April FOMC market with four outcomes — ">50 bps cut" trading at $0.001, "25 bps cut" at $0.008, "no change" at $0.985, "hike" at $0.001 — the asks sum to $0.995. A trader buying one share of each outcome pays $0.995 today and is guaranteed exactly $1.00 on resolution day, because the FOMC will land in exactly one of the four buckets. That is 50 basis points, risk-free, over whatever the time-to-resolution is, minus gas and fees. Polymarket's actual April 2026 Fed-decision market priced "no change" at roughly 99.9%, with the residual three outcomes splitting the rest, and traded a tight near-$1.00 sum precisely because professional desks sweep these residuals. The opportunity is not in the headline FOMC market in normal times; it is in the moments when one tail outcome's quote stales — when "hike" sits at $0.005 because nobody has updated it after a Powell speech — and the basket briefly drops below par.
Saguillo, Suárez-Tangil, Donet and Cuevas formalize this in Unravelling the Probabilistic Forest: Arbitrage in Prediction Markets (arXiv 2508.03474, August 2025), the first systematic on-chain study of overround/underround mispricings on Polymarket. Using historical order-book data, they identify two distinct strategies. The first is market-rebalancing arbitrage: complete-set buying or selling within a single multi-outcome market when the price simplex breaks par. The second is combinatorial arbitrage: trades across logically related but separately-listed markets where the same underlying state space is decomposed differently — for instance, a "Trump wins" binary versus a swing-state-by-swing-state decomposition versus an Electoral-College-bucket decomposition, all of which must reconcile by no-arbitrage. The headline empirical result: on-chain arbitrageurs realized profits near $40 million extracting these inconsistencies. The paper is also explicit that the opportunities are not random noise — they are structural, recurring, and concentrated in periods of low liquidity or high information arrival.
The historical analog on US-domiciled venues is PredictIt's 2014–2015 nomination markets. The Republican and Democratic nomination markets each listed dozens of candidate-YES contracts, and through 2014 and into mid-2015 the sum of all candidates' YES prices on the Republican side persistently exceeded $1.55 — a 55-cent overround. Researchers documenting the period (Arbitrage in Political Prediction Markets, JPM) found that the price sum compressed sharply around October 2015, after PredictIt introduced its "linked markets" mechanism (which let traders post collateral once across mutually exclusive contracts) and a 10% profit fee that reduced the after-fee economics of pure complete-set selling. That fee structure is itself instructive: PredictIt's overround was so large and so persistent that retail-scale arbitrageurs could fund a thesis for years off it, and the platform eventually had to intervene at the rules level. It is the prediction-market equivalent of decimalization compressing dealer spreads in 2001 — a structural change that destroyed an alpha source by altering the rules under which pricing happened.
The 2024 cycle on Polymarket showed both the maturity and the residual inefficiency. The headline Presidential Election Winner market traded $3.69 billion in cumulative volume across roughly 17 named candidates plus a residual "Other," and on election eve the sum of YES asks on the binary Trump-vs-Harris reduction was inside 99.9% of par. But earlier in 2024, with Biden, Newsom, and Michelle Obama all simultaneously priced in the 3–5% range while Harris and Trump anchored the top, the unconditional sum of all listed candidate YES prices wandered meaningfully above $1.00. The structural reason is well documented: retail flow concentrates on long-tail names — the "Michelle Obama steps in" narrative had $153 million of volume on its own, despite resolving at zero — and unless a maker is actively shorting those long-tail YES contracts, their bid stays elevated and the basket runs an overround. Per Dune-published Polymarket data, the Michelle Obama and Newsom contingent contracts traded with measurably wider spreads than Trump and Harris through summer 2024 — exactly the pattern Bürgi, Deng, and Whelan (CEPR DP 20631) identified across Kalshi: the favorite-longshot bias is the structural overround.
Their result deserves restatement because it is the cleanest empirical tie back to convertibles. Across more than 300,000 Kalshi contracts, contracts trading below 10 cents win materially less than their price implies — taker buyers of sub-10-cent contracts lose more than 60% of capital invested. Contracts trading above 90 cents win slightly more than their price implies and produce small positive expected returns. The bias is much stronger on the taker side than on the maker side: taker P&L averages roughly −32% across the contract universe, maker P&L roughly −10%. The implication is symmetric: long-tail YES is systematically overpriced, favorite YES is systematically underpriced (or fairly priced, depending on horizon). This is the same empirical signature as the 1990s convertibles market, where issuers and dealers underpriced the embedded long-vol option because their pricing apparatus was primitive and demand for the bond came from non-vol-sensitive buyers (insurers, traditional credit accounts). Retail PM traders are doing the reverse mistake — overpricing the long-tail option, underpricing the favorite — and for symmetric reasons (lottery preference, sentiment, narrative trading). Either way, the disciplined trade is to be the maker who sells the overpriced component and buys the underpriced one.
Translating this to executable strategy in late April 2026, three trade families fall out. First, complete-set sweeps in low-liquidity multi-outcome markets: scan Polymarket's Fed-decision per-meeting markets, ceasefire-by-date markets, and political-nominee markets for moments when the ask basket drops below $0.99 or the bid basket rises above $1.01. Saguillo et al. flag that these opportunities remain non-trivial in 2025 data, particularly during off-hours and immediately after news shocks before bots reprice the residuals. Second, combinatorial arbitrage across different decompositions of the same state space: the same election can be expressed as a binary, as a 50-state grid, as an Electoral-College-bucket grid, and as a popular-vote-margin grid — and these representations must reconcile. The arXiv 2603.03136 working paper on Polymarket microstructure explicitly invokes Shleifer-Vishny limits-to-arbitrage to explain why these don't always reconcile in real time. Third, the favorite-side maker book: post resting bids slightly inside spread on contracts trading above 90¢ in liquid markets where you have a calibration view, and harvest the 10-cent positive-EV residual that Bürgi, Deng, and Whelan identify. This is the closest direct analog to Griffin's 1990 trade — being a maker in the part of the price grid where the rest of the market is mispricing the option.
The structural ceiling on the strategy is real. Complete-set sweeps require capital in ratio to the number of outcomes (a 17-candidate market with one share of each at $0.06 average needs only ~$1.00 per complete set, but liquidity at the tail is thin and slippage destroys the trade above modest size). Polygon gas, settlement risk on UMA, and the bid-ask cost of entering all N legs simultaneously are the equivalents of convertible arb's financing costs and equity borrow fees — tractable, but they compress the gross overround into a net spread that determines whether the trade is worth doing. Combinatorial arb requires inventory across multiple markets simultaneously and is exposed to settlement-specification divergence in the same way Cardi B's Super Bowl halftime market resolved YES on Polymarket and NO on Kalshi for the same underlying event. Maker-side favorite trading is the most scalable but requires automated quote-pull on news monitoring, because adverse selection from informed flow — the Akey, Grégoire, Harvie, and Martineau pattern of 60-sigma-anomalous wallets — will eat the spread on any market where you are the slowest quote on screen.
Griffin's 1990 trade ended when convertible-arb capital grew from $768 million in 1994 to $25.6 billion by 2002 and returns compressed toward equity-vol carry alone. The same arc is unambiguously underway on prediction markets: Saguillo et al.'s $40 million realized arbitrage profit is small relative to the $14 billion of cumulative volume that produced it, and the per-trade gross is shrinking as Jump, Susquehanna, and now-funded Polymarket-native shops sweep residuals. But the trade is still there. The favorite-longshot bias is a structural overround on retail flow that has persisted across every well-studied betting market for a hundred years — Snowberg and Wolfers in horse racing, Bürgi-Deng-Whelan on Kalshi event contracts, Le 2026 across Polymarket politics. As long as retail traders treat 5-cent contracts as lottery tickets and 90-cent contracts as boring, the maker who supplies the other side captures the overround. That is the 1990 convertibles trade, executed in 2026 against a different population of mispricers.