Start with the convergence mechanism
The word arbitrage can describe identical-asset price differences, relative-value relationships, futures basis, funding dislocations, triangular conversion loops, or onchain routes. These structures do not share the same guarantee of convergence. An identical spot asset may be transferable in principle but operationally trapped; a statistical pair may never converge; a dated future has an expiry but introduces margin and venue risk.
A useful strategy specification states the instruments, venues, sides, position size, expected holding interval, conversion or settlement path, and residual exposure. It also states what evidence invalidates the trade. This prevents a model from changing its story after a loss—for example, calling a failed arbitrage a long-term investment.
Six strategy families
- Cross-venue spot: buy and sell the same economic asset across pre-funded venues.
- Triangular conversion: complete a currency loop within one venue or settlement environment.
- Pair trading: trade a modeled spread between related assets using a hedge ratio.
- Statistical arbitrage: form a portfolio from cross-sectional or time-series forecasts.
- Basis and carry: trade spot against futures or perpetual contracts while modeling financing and margin.
- Onchain arbitrage: compose swaps, loans, and settlement instructions across smart contracts.
Low-latency execution can be applied to several families, but speed is not a separate convergence mechanism. It changes the probability of reaching a quote and the severity of operational error. Agentic AI is similarly an orchestration layer: it can assist selection and monitoring without changing the economic structure of the underlying trade.
Compare routes by required resources
Cross-venue strategies consume inventory and counterparty capacity. Pair trading consumes borrow or derivatives capacity and depends on relationship stability. DeFi routes consume blockspace, smart-contract trust, and transaction-construction reliability. Basis trades consume collateral and financing. A strategy that appears capital-light in one dimension can be heavy in another.
Build a resource table before backtesting. Include data feeds, venue accounts, approved instruments, balances, credit, wallets, gas assets, borrow access, legal permissions, monitoring, and incident ownership. A route is not deployable when one critical resource exists only in a spreadsheet assumption.
Use executable, not decorative, prices
Every strategy should calculate the prices available for the intended quantity. That means walking order-book depth, applying pool curves, or modeling impact. Use the correct side of the market for entry and exit. Add fee tier, funding, borrow, gas, conversion, rebalancing, and a reserve for uncertainty. Then stress the inputs.
A robust threshold varies with liquidity, volatility, data health, and inventory. Fixed percentage alerts can be useful for observation but should not authorize execution. When the conservative edge is negative or uncertain, the correct strategy output is no trade.
Progress from research to bounded operation
A staged path begins with a written hypothesis and historical data audit. Replay tests follow, then live shadow mode, where the system sees current data but cannot place orders. Supervised execution with small limits can test venue semantics and reconciliation. Automatic operation, where appropriate, should remain bounded by independent controls and reversible permissions.
Promotion criteria should include data quality, model stability, realized cost error, incident recovery, and abstention quality—not just cumulative return. Retirement criteria are equally important. A strategy can be technically healthy and still lose its economic edge as participants, fees, or market structure change.