Arbitrage

The mechanics of arbitrage

Suppose Apple trades at $200.00 on Nasdaq and $200.05 on a smaller exchange at the same instant. An arbitrageur buys Apple at $200.00 and sells at $200.05 simultaneously, locking in $0.05 per share regardless of where the price moves next. The trade has zero net market exposure — the long and short cancel — yet generates positive cash flow. Repeat ten million times and the arbitrageur captures $500,000 risk-free. Of course the spread closes the moment they trade.

The textbook condition is the no-arbitrage principle: in a frictionless market with rational participants, no portfolio can have zero cost, non-negative payoffs in every state, and a positive payoff in some state. If such a portfolio existed, traders would scale it infinitely, instantly bidding away the discrepancy.

Taxonomy of arbitrage

Type	Example	Risk	Typical practitioner
Pure (spatial)	Same stock, two exchanges	Execution risk only	HFT market makers
Triangular FX	EUR/USD × USD/JPY ≠ EUR/JPY	Microsecond execution	FX banks, HFT
Cash-and-carry	Spot oil + storage − futures	Storage cost, basis risk	Commodity merchants
ETF creation/redemption	ETF price ≠ NAV of basket	Authorized-participant risk	Authorized participants (Citadel, Susquehanna)
Convergence (relative value)	Off-the-run vs on-the-run Treasuries	Spread divergence, funding	Hedge funds (LTCM, Citadel)
Merger / risk arbitrage	Long target, short acquirer in stock deal	Deal break risk	Event-driven funds
Statistical arbitrage	Long-short pairs based on history	Model risk, regime change	Quant funds (Renaissance, Two Sigma)
Latency arbitrage	Slow venue lags fast venue	Infrastructure cost	HFT (Virtu, Jump)

Worked example: LTCM's convergence trade, August 1998

LTCM noticed that the 30-year Treasury issued in February (off-the-run) yielded ~12 basis points more than the brand-new 30-year issued in August (on-the-run). Both pay the same government and have nearly identical duration; the spread reflected liquidity premium for the newest bond. Historical data showed the spread mean-reverted to ~5 bp within months as the on-the-run aged.

The trade: buy the off-the-run, short the on-the-run, hold to convergence. Each $1 billion of long earned roughly 12 bp × $1 billion = $1.2 million per year if held; convergence to 5 bp would unlock another $7 million on the remaining duration. Tiny in absolute terms — so LTCM levered 25:1 via repo financing, turning $5 billion of equity into $125 billion of positions.

In August 1998, Russia defaulted on its debt. Investors fled to liquidity, dumping older Treasuries and crowding into the newest on-the-run. The spread widened from 12 bp to over 20 bp in days. LTCM's positions showed mark-to-market losses; lenders demanded more collateral; LTCM had to sell into a falling market, widening the spread further. By late September the fund had lost $4.6 billion. The Fed organized a $3.625 billion private-sector bailout from 14 banks to prevent forced liquidation. The spread did eventually converge — by 1999 — but LTCM was already wound down.

Lesson: a convergence trade is correct on average and ruinous in tail events. Markets can stay irrational longer than you can stay solvent (Keynes, attributed). Shleifer and Vishny's 1997 paper "The Limits of Arbitrage" formalized this: arbitrageurs face capital constraints exactly when mispricings are largest.

Arbitrage vs related concepts

	Pure arbitrage	Risk arbitrage	Speculation	Hedging
Net market exposure	Zero	~Zero	Long or short	Offsets existing exposure
Profit certainty	Locked in	Probabilistic	None	Reduces uncertainty
Capital required	Minimal (margin)	Variable	Full position	Cost of hedge
Time horizon	Microseconds–days	Days–months	Variable	Matched to exposure
Worst case	Execution loss	Spread divergence	Total loss	Hedge cost
Practitioners	HFT, market makers	Hedge funds	Retail, prop, institutions	Corporates, airlines, farmers

From medieval bills of exchange to microwave HFT

Arbitrage predates formal finance. Medieval Florentine merchants issued bills of exchange across Lyon, Bruges, and Genoa, profiting on regional price differences in gold and silver. Nathan Mayer Rothschild allegedly profited on news of Waterloo via carrier pigeon — arguably an information arbitrage. The 1602 Amsterdam Stock Exchange traded VOC shares against Lyon, with merchants riding boats to bridge the price gap.

The modern era began with the 1980s rise of stat-arb at Morgan Stanley and the launch of futures-cash basis trading at the CME. Latency arbitrage became dominant after 2007 when Reg NMS fragmented US equity markets into 13 venues; the Spread Networks fiber line from New Jersey to Chicago (completed 2010) shaved 3 milliseconds and reportedly cost $300 million. By 2013 microwave links replaced fiber on that route, shaving another 4 milliseconds. As of 2024, top firms operate laser links above clouds and co-locate within meters of matching engines.

Variants and frontiers

• Statistical arbitrage. Pioneered by Bamberger and Tartaglia at Morgan Stanley (1985), perfected by Renaissance Medallion (1988-present, ~66% gross annualized). Uses pair trading, factor models, mean-reversion signals across thousands of equities.
• Merger arbitrage. Long the target at announced deal price minus current price; short the acquirer if a stock-for-stock deal. Risk: deal breaks. Spreads embed implied break probability.
• Capital-structure arbitrage. Long bonds, short equity (or vice versa) when the Merton structural model says they're inconsistently priced.
• ETF arbitrage. When SPY trades above NAV of its 500 stocks, an authorized participant delivers the basket to BlackRock for new ETF shares and sells those shares — flattening the gap.
• Crypto arbitrage. Cross-exchange (BTC on Binance vs Coinbase), cross-chain (wrapped vs native), and triangular within DEX pools. Heavily pursued in 2017-2021; spreads have compressed substantially.
• Index-rebalance arbitrage. Front-run forced buying when stocks are added to S&P 500. The "S&P effect" — added stocks gain ~5% on inclusion — has been exploited and partially eroded since the 1990s.

Famous real-world cases

• 1990s — Salomon Brothers' arbitrage desk. Run by John Meriwether (later LTCM founder); pioneered Treasury convergence trades and bond-options relative value. Profitable until 1991 Treasury bidding scandal.
• 1998 — LTCM collapse. $4.6 billion loss, Fed-organized bailout. Documented in Roger Lowenstein's When Genius Failed.
• 2008 — quant earthquake (August). Multiple stat-arb funds delevered simultaneously; correlated unwind drove single-day losses of 20%+ at Renaissance, AQR, Goldman GEC.
• 2010 — flash crash. ETF-NAV arbitrage broke briefly; SPY traded at 50% of NAV before circuit breakers fired.
• 2020 — Treasury basis trade. $800 billion of hedge-fund leverage on cash-futures basis blew up in March 2020; the Fed had to backstop Treasury markets to prevent a doom loop.
• 2024 — Japanese yen carry unwind. Borrow yen at near-zero, invest in higher-yielding currencies. The August 2024 unwind erased weeks of gains in days when BOJ raised rates 25 bp.

Common pitfalls

• Ignoring funding cost. Levered arbitrage requires repo or prime-broker financing; in a crisis, haircuts widen and lenders demand more collateral exactly when you can't supply it.
• Underestimating crowdedness. If everyone runs the same convergence trade, the unwind is correlated. The August 2007 quant crisis showed that.
• Ignoring transaction costs and taxes. A 5 basis point spread in a 0.30% commission market is negative carry.
• Conflating arbitrage with risk-free. All real arbitrage has execution, counterparty, and funding risk. Pure no-arbitrage is a textbook abstraction.
• Trading too small to overcome costs. HFT arbitrage requires microsecond infrastructure; retail is structurally late.
• Overestimating mean-reversion. "The spread always normalizes" — until it doesn't. Position sizing and stop-losses matter.