On September 28, 1981, Kohlberg Kravis Roberts agreed in principle to take Arcata Corporation private at $37 a share. Two days later, Warren Buffett began buying. Over the next eight weeks Berkshire accumulated roughly 400,000 shares — about 5% of the company — at an average price near $33.50. When the deal closed Berkshire booked roughly $1.7 million on the cash leg, an annualized return in the 15% range on a position that functioned, until settlement, as something close to a Treasury substitute. The interesting part of the trade was a piece of paper KKR had written into the merger agreement: in 1978 the federal government had condemned 10,700 acres of Arcata's old-growth redwood for $97.9 million to expand Redwood National Park, and Arcata was disputing both the price and the interest rate. KKR offered shareholders two-thirds of any additional sum the courts eventually awarded. In 1988, that residual claim settled for $519 million. Berkshire collected an extra $19.3 million on a position it had nominally exited six years earlier. Buffett's 1988 letter would disclose that the firm had earned $78 million on $147 million of average invested arbitrage capital that year — a 53% pre-tax return.
This is the trade Polymarket reinvented in 2024.
The classical mechanic
Risk arbitrage as an institutional discipline starts with Gustave Levy. Levy joined Goldman Sachs in 1933 as a foreign-bond trader, made partner in 1945, and through the 1940s built what the firm now describes as one of the first risk-arbitrage desks on Wall Street. The desk that Robert Rubin would inherit and that Ivan Boesky would imitate at scale was Levy's. The mathematical structure has not changed in the eighty years since: a deal is announced at price P, the target trades at price T < P, the arbitrageur captures (P − T) over the closing window. A 5% spread that closes in three months is a 20% annualized return; a 1% spread that closes in two weeks annualizes to roughly 26%. The strategy is uncorrelated with broad equity beta, scalable into the hundreds of millions, and dominated by two specific risks: deal break (the merger fails and the target reverts to its undisturbed price) and timing extension (the spread compounds over a longer denominator).
Boesky industrialized the playbook through the early 1980s, deploying around $200 million of capital at peak. The strategy collapsed when SEC enforcement worked backward from Drexel Burnham banker Dennis Levine, who had been selling Boesky nonpublic merger tips. On November 14, 1986, the SEC's settlement was made public: Boesky agreed to a $100 million disgorgement-and-penalty payment, a permanent industry ban, and what would become a 22-month prison sentence. His cooperation produced Michael Milken; the broader scandal produced the Insider Trading and Securities Fraud Enforcement Act of November 19, 1988, which raised statutory penalties, made firms liable for failing to police employees, and authorized cash bounties to informants. The structural lesson — that a strategy whose edge depends on short windows and tight spreads rewards information leakage with criminal precision — is the part of the Levy/Boesky/Buffett legacy that has ported most cleanly to Polymarket.
Tail-end trading: the same trade, sixty years later
The Polymarket equivalent is buying YES contracts at $0.95 to $0.99 after the underlying event has factually occurred but before on-chain settlement. The trader nicknamed "Fish" told BlockBeats that roughly 90% of large Polymarket trades over $10,000 execute above $0.95, with arbitrageurs supplying liquidity to retail sellers who would rather have free capital for the next bet than wait out settlement. Fish's framing is exactly Levy's: "Although each trade only earns 0.1% profit, if the capital is large enough and the frequency is high enough, it can accumulate into a considerable income."
The annualized math is more attractive than 1980s merger arb because the windows are shorter. Polymarket's UMA Optimistic Oracle v2 normally resolves a market within a 2-hour challenge window after a $750 USDC bond is posted by a whitelisted proposer; if undisputed the market settles immediately. A disputed proposal escalates to UMA's Data Verification Mechanism, where token holders vote over a 48-to-96-hour window — call it 4 to 6 days end-to-end. Kalshi, a CFTC-regulated Designated Contract Market, typically settles 1 to 12 hours after market close once the source-agency data arrives, plus a settlement leg of about 3 hours. A 1% spread captured in a 6-hour window annualizes to north of 1,400% if the cycle can be repeated, before inventory and capital constraints. In practice, professional Polymarket tail-end desks compound much smaller realized returns over many concurrent positions, which is precisely the structure of a 1980s merger-arb book.
The directional accuracy of Polymarket prices in the run-up to settlement is the modern analog of the merger-spread tightening curve. Aggregate platform-wide statistics published on Dune show that the leading outcome at four hours, twelve hours, and one day before settlement matches the eventual resolution roughly 95%, 89%, and 88% of the time on resolved markets. That is not "wisdom of crowds." It is the same convergence pattern that allowed Buffett to claim a 39% IRR on the Arcata trade — prices tighten as the residual uncertainty collapses.
Deal-break risk, UMA edition
Classical merger arb's nightmare is the regulatory rejection or financing failure that breaks the deal at $0.10 below the bid. The prediction-market analog is resolution dispute risk, and several documented 2024–2026 cases have produced total wipeouts on positions held above $0.95.
The Zelenskyy "suit" market — a $237 million volume contract on whether Volodymyr Zelenskyy would be photographed in a suit between March 22 and June 30, 2025 — resolved NO on July 1, 2025. UMA voters cited insufficient "credible reporting consensus" on Zelenskyy's June 24 NATO summit appearance in the Hague, where he wore a black collared jacket and matching trousers that several major outlets called a suit and others did not. The dispute exposed a structural vulnerability: UMA's token market capitalization was roughly $95 million while a single market commanded $242 million of volume, and the top ten UMA voters held about 30% of average voting weight per dispute. A trader holding YES at $0.97 the day before resolution lost everything.
The TikTok ban market ($120 million volume) resolved YES on January 20, 2025, hours before President Trump signed an executive order delaying enforcement of the Protecting Americans from Foreign Adversary Controlled Applications Act for 75 days. The technical case for YES — that the law had taken effect on January 19 — was defensible; the spirit-of-the-question case for NO was just as defensible. The DVM was bypassed and the market resolved direct. The Cardi B Super Bowl LX halftime contract on February 8, 2026 split between platforms: Polymarket ($10 million volume) ultimately resolved YES on the credible-reporting-consensus standard; Kalshi ($47.3 million volume) paused its market and settled at the last traded price of $0.74 NO / $0.26 YES under its stricter "performing" specification. Cross-platform arbitrage legs that had been built on the assumption of a common settlement specification went to zero on at least one side.
These are the "deal breaks" of the Polymarket era. Traders buying at $0.95 are pricing a 5% probability of resolution dispute risk — historically a generous spread on uncontested events and a guaranteed wipeout on the wrong side of an ambiguous one.
What the data says about the trade
The most rigorous empirical work on prediction-market profitability is Pat Akey, Vincent Grégoire, Nicolas Harvie, and Charles Martineau's "Who Wins and Who Loses in Prediction Markets? Evidence from Polymarket," dated March 18, 2026. The authors analyze more than 1.4 million users across 70 million trades and over $20 billion in volume from 2022 through 2025. Three findings matter for tail-end trading specifically. First, the top 1% of users capture 84% of all trading gains, and 70.8% of users are net losers — the platform structure rewards the disciplined arbitrageur and punishes the lottery-ticket buyer. Second, market-making (resting limit orders rather than crossing the spread) is the single largest predictor of profitability. Third, in their companion Harvard Law School Forum piece, the same authors construct a five-signal screen across roughly 93,000 markets and 50,000 wallets that flags traders with a 69.9% win rate, exceeding the permutation null by over 60 standard deviations and accounting for an estimated $143 million in anomalous aggregate profits.
The "Anatomy of Polymarket" working paper (arXiv 2603.03136) by Tsang and Yang adds the microstructure piece. Using full Polygon transaction-level data for the 2024 election cycle, the authors document that arbitrage deviations narrowed and Kyle's lambda — the price impact per unit of order flow — declined by more than an order of magnitude as the market matured between January and November 2024. The decline in lambda is the prediction-market analog of equity dealer-spread compression in the late 1990s. Crucially, the same paper develops a volume decomposition that separates exchange-equivalent trading volume from share-minting, burning, and conversion flows. Naive aggregation of on-chain data overstates real economic trading by a wide margin, which is why prior estimates of Polymarket arbitrage profitability were systematically too high.
The cleanest direct estimate of arbitrage capture sits in Saguillo, Ghafouri, Kiffer, and Suárez-Tangil's "Unravelling the Probabilistic Forest" (arXiv 2508.03474, August 2025). The IMDEA Networks team analyzed 86 million bets across thousands of Polymarket markets from April 2024 through April 2025 and identified roughly $40 million of realized arbitrageur profits across two distinct strategies: market-rebalancing arbitrage within single multi-outcome markets, and combinatorial arbitrage across logically related markets. Most individual instances generated 1% to 5% trade-level returns — exactly the magnitude Fish described to BlockBeats. The top three wallets placed more than 10,200 bets and netted approximately $4.2 million.
Current markets and the live spread
As of late April 2026, Polymarket's "Fed decision in April?" market has accumulated $183.9 million in volume since its November 13, 2025 launch and prices the no-change outcome at roughly $0.999 — a textbook tail-end market with two trading days remaining before the April 28–29 FOMC meeting. The "Bitcoin all time high by ___?" event ($6.3 million volume since December 2025) and "What price will Bitcoin hit in 2026?" ($33.3 million) settle off Binance 1-minute BTC/USDT high prices, which means the resolution function is mechanical and the dispute surface is small — a structurally cleaner tail-end venue than the Zelenskyy-suit class of markets. The Khamenei-out-as-Supreme-Leader complex (multiple monthly expirations, $49.7 million volume on the January contract and $131.1 million on the February contract before the February 2026 US-Israeli strike that resolved much of the tree) shows the opposite extreme: dense political content, stacked expirations, and the kind of ambiguity that produced the $237 million Zelenskyy dispute.
The tail-end discipline is to choose markets where the resolution function is mechanical and verifiable, the underlying event has factually occurred, and the on-chain settlement gap is the only remaining variable. The Iran ceasefire cycles of June 2025 and April 2026 are the cautionary opposite: at least 50 newly created Polymarket accounts placed concentrated YES bets in the hours and minutes before President Trump's Truth Social ceasefire announcement on the April 2026 cycle, with Bubblemaps clustering roughly $600,000 of profits to a coordinated wallet ring across the $170 million total volume. Senators Schiff and Blumenthal demanded CFTC investigation; Representatives Torres and Moore filed bipartisan legislation. The April 25, 2026 DOJ indictment of U.S. Special Forces soldier Gannon Ken Van Dyke for using classified Maduro-operation intelligence to bet $33,000 into roughly $400,000–$440,000 on Polymarket is the first criminal prosecution for prediction-market insider trading in U.S. history. It will not be the last.
What the structural risk actually is
The Boesky case in 1986 produced the Insider Trading and Securities Fraud Enforcement Act of 1988 within twenty-four months. The Van Dyke case in April 2026 will move on a similar timeline. The strategic question for a tail-end trader is not whether prediction markets become more regulated — they will — but whether the structural edge survives the regulation. Three observations.
First, the favorite-longshot bias documented by Bürgi, Deng & Whelan (CEPR DP 20631) and Le (2026, arXiv 2602.19520) implies that high-price favorites are systematically underpriced relative to true probability while longshots are overpriced. This is the structural source of tail-end alpha and it is invariant to enforcement regime. A regulated Polymarket and a regulated Kalshi will both display favorite-longshot bias as long as retail flow dominates the take-side of the order book.
Second, the resolution-dispute risk premium of 1.5% to 4.5% that persists on identical contracts across Polymarket and Kalshi is a real cost of capital, not a market inefficiency. The Cardi B and government-shutdown cases prove the spread cannot be fully arbitraged because settlement specifications can and do diverge. A tail-end trader who buys at $0.97 on Polymarket against a $0.94 short on Kalshi is short the joint distribution of two settlement processes, not long a single-leg arbitrage.
Third, Akey et al.'s finding that moving from pure taker to pure maker reduces the probability of losing money by approximately 36 percentage points is the most economically significant statistic in the empirical literature. Tail-end trading executed as a maker — resting bids at $0.96, $0.97, $0.98 against retail sells — is structurally a different trade from tail-end trading executed as a taker lifting offers at $0.99. The former is Levy's desk; the latter is Boesky's tip-driven leverage stack without the tips. The 1990s arc of equity-market dealer rents — Christie-Schultz collusion paper in 1994, order-handling reforms in 1997, decimalization in 2001 — is the template for what will happen to Polymarket's spread structure over the next five years. Tail-end traders who build the maker-side infrastructure now will compound; tail-end traders who lift offers and pay for the privilege will not.
Buffett's Arcata trade earned 15% annualized on the cash-equivalent leg and a 6-year-deferred bonus on the redwood claim. The redwood claim was a free option attached to a deal that had already happened. The Polymarket tail-end trade is the same option, repriced every two hours, settled on-chain, and exposed to a dispute mechanism that did not exist in 1981. The mechanic that Levy formalized and Buffett extended is operational on Polymarket today. The traders who recognize this — and who price the UMA dispute risk correctly — will earn the spread. The traders who confuse a 95-cent contract for a riskless one will fund them.