Turning a Research Signal Into a Live Strategy

A signal that survives a backtest still needs an execution layer, a risk framework, and a monitoring plan before it earns a dollar of live capital.

Turning a research signal into a live strategy requires building an execution layer, a risk framework, and a monitoring plan around a result that, on its own, is only a statement about historical data. A signal that clears every backtest check described elsewhere on this site, including a proper out of sample split and a realistic cost model, still fails in live trading if the infrastructure around it treats the transition from research to production as an afterthought rather than a distinct engineering problem.

Before any capital is committed, a strategy should run in a paper trading environment that receives live market data and generates real orders, but routes those orders to a simulated fill engine rather than an actual exchange. This step exists to catch a category of bug that a historical backtest cannot surface: data feed outages, unexpected message formats, and timing assumptions in the code that only break when the strategy is running against a live, asynchronous data stream rather than a neatly ordered historical file. A strategy that fails during paper trading has usually found an infrastructure bug rather than a flaw in the underlying signal, and fixing that bug before risking capital…

Once paper trading is running, the fills it generates should be compared directly against what the backtest predicted for the same trades, using the same instruments over the same period. A persistent gap, where paper trading shows meaningfully worse fill prices than the backtest assumed, indicates the backtest's cost model, discussed in Transaction Cost Modelling Inside a Backtest, understated real execution cost, and that gap should be fed back into the backtest's assumptions before capital sizing is finalized, not discovered for the first time after real money is at risk.

A strategy's first live allocation is commonly sized well below the level its backtest and paper trading results would nominally support, often to a small fraction of eventual target size, specifically to limit the cost of discovering a problem that only appears with real capital and real counterparties, such as a broker's handling of partial fills or an unexpected interaction with existing positions elsewhere in the book. This initial period functions as a final validation stage, and the strategy is scaled up toward its full target size only after a defined number of trading days or a defined number of trades has passed without a material deviation from expected performance.

Every live strategy needs an automated kill switch, a rule that halts new order submission and, where appropriate, flattens existing positions, triggered by conditions defined before the strategy went live rather than judged in real time under pressure. Common triggers include a drawdown exceeding a fixed threshold over a fixed lookback window, a data feed going stale beyond an acceptable delay, or the strategy's realized trade count in a session falling well outside its historical range in either direction, since an unusually low trade count can indicate a broken signal path just as clearly as an unusually high one can indicate a runaway order loop.