In 1977, John Meriwether started a small group inside Salomon Brothers whose only job was to find pricing inconsistencies between related fixed-income instruments and trade them until they vanished. The desk was institutionally awkward — a quant cell embedded in a relationship-driven bond house — and it tolerated the awkwardness because it produced money. By the mid-1980s, by the accounts captured in Michael Lewis's Liar's Poker (1989), Meriwether's arbitrage group was reportedly responsible for somewhere on the order of 80 to 100 percent of Salomon's earnings. Its core trades — on-the-run versus off-the-run Treasuries, swap spreads, yield-curve flatteners — were not directional bets on rates. They were bets on the shape of a curve, on the proposition that two instruments tied by an arithmetic identity must, at expiry, agree.

Meriwether took the playbook private in February 1994. Long-Term Capital Management opened with $1.25 billion of investor capital and a board that included Myron Scholes and Robert C. Merton, who would share the 1997 Nobel Memorial Prize in Economics for the Black-Scholes model. The fund returned roughly 21 percent in 1994, 43 percent in 1995, and 41 percent in 1996, net of fees, on leverage that publicly available accounts (including the Federal Reserve Bank of Richmond's Region Focus and Roger Lowenstein's When Genius Failed) place between 25:1 and 30:1 of debt to equity. The mechanism was the same as Salomon's: identify a curve whose internal arithmetic was off, take offsetting positions, and wait for convergence.

What killed LTCM in August 1998 is also what makes the strategy worth porting now. Russia's August 17, 1998 default on roughly $13.5 billion of ruble-denominated debt produced a global flight to the most liquid benchmark instruments. The relative-value spreads LTCM was short widened against the off-the-run, less-liquid securities it was long. Correlations across what had been treated as independent trades — sovereign convergence, on/off-the-run, swap spreads — jumped from a stress-tested 0.3 to roughly 0.7. Every convergence trade lost simultaneously. The lesson is not that the underlying arithmetic was wrong; the arithmetic held, and the spreads eventually converged. The lesson is that the shape of a term structure can be locally inconsistent while remaining globally exposed to a single liquidity factor.

That intellectual core — a term structure of related instruments contains an embedded forward curve, inconsistencies are tradable, and the residual risk is a flight-to-liquidity factor that touches every leg at once — ports almost without modification to today's prediction markets.

The hazard-rate curve embedded in nested ceasefire markets

Polymarket runs nested binary markets on whether a Russia-Ukraine ceasefire will be reached by a given date. As of late April 2026, the visible term structure (verified against the live event pages) is approximately:

by April 30, 2026: 0.5 percent
by May 31, 2026: 5.2 percent
by June 30, 2026: 7.5 percent
by December 31, 2026: 25.5 percent

Each market is a separate contract with its own UMA resolution spec, but logically they are nested: the event "ceasefire by April 30" is contained in "ceasefire by May 31," which is contained in "ceasefire by June 30," which is contained in "ceasefire by December 31." The first no-arbitrage requirement is monotonicity: P(T₁) ≤ P(T₂) for T₁ < T₂. If May 31 ever printed below April 30, you could buy May, sell April, and pocket the difference at resolution regardless of outcome. The current 0.5 / 5.2 / 7.5 / 25.5 strip respects monotonicity, so the calendar arbitrage is not present in levels.

The interesting object is the marginal hazard rate — the conditional probability that the ceasefire occurs in interval (Tᵢ, Tᵢ₊₁] given it had not occurred by Tᵢ. Bootstrapping from the survival probabilities S(T) = 1 − P(ceasefire by T):

S(Apr 30) = 0.995, S(May 31) = 0.948, S(Jun 30) = 0.925, S(Dec 31) = 0.745
h(Apr→May) = 1 − 0.948/0.995 ≈ 4.7%
h(May→Jun) = 1 − 0.925/0.948 ≈ 2.4%
h(Jun→Dec) = 1 − 0.745/0.925 ≈ 19.5% over six months, ≈ 3.6% per month if you smear it linearly

The shape is non-monotonic. The market is paying for a relatively elevated April-to-May hazard, then a near-term lull, then a gradual rise into year-end. A 1990s credit-arb desk would recognize this profile immediately: it is the discrete analog of a humped CDS curve, where short-dated default risk is being priced higher than the next interval. The economic interpretation is a known catalyst — talks, summit, deadline — concentrating probability mass into a specific window. The trading question is whether the kink is correct.

Three things follow from the kink. First, the implied schedule of catalysts is testable against news. If Polymarket says April-to-May is 4.7 percent and the only April catalyst on the calendar is a low-information-content meeting, the market is over-paying for the near term. Second, the curve admits structured trades: long the May contract, short two units of the June contract is roughly delta-neutral to the April catalyst but pays off if the May-to-June hazard mean-reverts upward. Third, the curve must reconcile with any longer-dated parent. The "by end-2026" market at 25.5 percent must equal one minus the product of monthly survival probabilities back to today; if the monthly hazards bootstrapped from sub-markets do not multiply out to that number, one of the legs is wrong.

The three-leg combinatorial check on Fed-decision markets

The same arithmetic applies to Polymarket's Fed-decision strip, and here the no-arbitrage check has three independent legs that must reconcile, not two. As of late April 2026 the visible per-meeting "no change" probabilities are approximately 99.7 percent for April 28-29, 93.5 percent for June 16-17, and 85.5 percent for July 28-29. Polymarket also runs "first cut by date X" markets — "first cut by October" near 54 percent, "first cut by December" near 60 percent — and a multi-outcome "how many cuts in 2026" distribution where the 40.1 percent mass on zero cuts is one bar of a complete probability simplex.

Three logically independent legs:

Per-meeting decision markets. The probability of "no cut at any of the next k meetings" is, under the (strong but standard) independence assumption, the product of the per-meeting "no change" probabilities. April × June × July ≈ 0.997 × 0.935 × 0.855 ≈ 0.797. So per-meeting markets imply that the probability of no cut by the end of July is about 79.7 percent, i.e., probability of first cut by the July meeting or earlier is about 20.3 percent.
First-cut-by-date markets. A "first cut by July" contract should print near 0.20 to be consistent with leg 1. A "first cut by October" contract at 0.54 implies the August-to-October window absorbs roughly 0.34 of probability mass; given there is one meeting in that window (September), this implies a September "no change" probability around 1 − 0.34/0.797 ≈ 57 percent. If Polymarket's individual September-decision contract prints far above that, the inter-market spread is the trade.
Cumulative-count markets. "How many cuts in 2026" is a discrete distribution over {0, 1, 2, 3, 4+}. The probability of zero cuts must equal one minus the probability of "first cut by the last 2026 meeting" computed from leg 2 (about 1 − 0.60 = 0.40, against the observed 0.401 — consistent, within the rounding the orderbook will show). The mean of the distribution must equal the implied path average from per-meeting expected cut sizes. The variance must be consistent with the dispersion of "first cut by X" prices across the year.

These three legs are the prediction-market analog of the relative-value triangle Salomon's group and LTCM ran on swap spreads, on/off-the-run Treasuries, and Eurodollar futures: three windows on the same forward curve, and any persistent disagreement is a tradable signal. The empirical question is how often the triangle is closed. Anecdotally, the per-meeting markets and the count market track each other within a percentage point or two on resolution-eve; "first cut by date" markets are visibly noisier, especially at intermediate dates where there is no nearby orderbook anchor. That noise is the alpha.

The formal treatment

The most direct theoretical work on this object is Shaw Dalen's "Toward Black-Scholes for Prediction Markets: A Unified Kernel and Market Maker's Handbook" (arXiv:2510.15205), submitted October 17, 2025 and revised April 6, 2026. Dalen treats traded probabilities as risk-neutral martingales under a logit jump-diffusion kernel, separates diffusive belief volatility from jump intensity via expectation-maximization, and constructs what he calls "calendar variance swaps" and "threshold notes" as derivatives written on the resulting belief-volatility surface. The framework formalizes the intuition that a nested term structure of binaries is a forward curve in disguise, and that the curve admits a derivatives layer once the kernel is specified. As of April 2026 this is the most theoretically rich and least empirically explored alpha pocket in prediction markets — the academic scaffolding exists, but the empirical density of identifiable inconsistency, the persistence half-life, and the per-trade Sharpe at retail size are essentially uncatalogued.

The structural risk is exactly LTCM's

LTCM blew up not because its convergence trades were wrong on the arithmetic but because the residual risk factor — the price of liquidity — moved against every leg simultaneously. The prediction-market analog is structurally identical and arguably tighter in time. A single resolution event — a real ceasefire, a Fed pivot announcement, a major geopolitical shock — can reset the entire term structure of related markets in a single block. If you are long a kink and short the wings as a relative-value bet, you are implicitly assuming the catalyst either fires inside your window or the curve smoothly re-prices around your trade. A discontinuous resolution — a ceasefire announced on a Tuesday morning — collapses every nested contract on the curve to either 1 or 0, and your "neutral" combinatorial position is suddenly directional.

A second risk has no clean LTCM analog: settlement-spec divergence. Polymarket's UMA Optimistic Oracle and Kalshi's CFTC-registered settlement committee can and do disagree on what counts as the same underlying event. The Cardi B Super Bowl halftime contract resolved YES on Polymarket and NO on Kalshi. A government-shutdown contract did the same. For a calendar-spread trade run within a single venue this is not a problem. For a cross-venue convergence trade — long Kalshi, short Polymarket on what looks like the same event — the resolution itself can fail to converge, which is correlation breakdown of a kind 1990s fixed-income arbitrageurs never had to model.

A third, smaller risk: UMA dispute extension. The standard 2-hour challenge window with $750 USDC bond can extend to a 4-6 day Data Verification Mechanism escalation if challenged. A combinatorial position that depends on simultaneous resolution can find one leg locked while the other prints, opening a window in which the "completed" leg is freely tradable against a frozen counter-leg.

The retail-versus-institutional ceiling

The arithmetic of the curve is freely available: every Polymarket event page exposes a bid-ask, every nested market is enumerable through the public API, and the bootstrap requires no proprietary data. What retail cannot do is execute the multi-leg trade at the size and speed required to capture small inconsistencies. A 30-basis-point kink in a hazard-rate curve, repeated across a dozen ceasefire and Fed contracts, is a serious annualized return at $50M of capital and an irrelevance at $5,000. Spread-capture on calendar arbitrage is a market-maker game; relative-value on hazard-rate kinks is a directional game with quantitative scaffolding. The Susquehanna-Kalshi liquidity relationship and Jump Trading's February 2026 equity stakes in both Kalshi and Polymarket suggest the institutional bid for exactly this strategy is being assembled now.

The current markets a reader can monitor — in roughly increasing order of analytical density — are Polymarket's nested ceasefire strip on Russia-Ukraine for 2026 (April through December), the Fed-decision per-meeting strip alongside the "first cut by X" and "how many cuts in 2026" markets, and any nested geopolitical curves that emerge around defined deadlines (debt-ceiling, tariff-extension, election-cycle markets). The discipline is to bootstrap the implied hazard rate from each, check the three-leg combinatorial reconciliation, identify which leg is the noisy one, and size the trade so that a discontinuous resolution event — the prediction-market analog of August 17, 1998 — does not take the book down with it. Meriwether's group ran that arithmetic for two decades before the residual risk found them. The same window is open on Polymarket today, and the curves are visibly less efficient than 1980s Treasuries were when the Salomon arb group started taking them apart.