Execution quality is the silent alpha source

Over the years we have killed more strategies for execution reasons than for model decay. The model was correct, the signal was real, the backtest was honest. Then the strategy went live and the spread, slippage, and rejection profile of the venue ate the edge before any of it reached the equity curve. This is the boring failure mode that nobody warns about loudly enough.

The asymmetry is harsh. A small modelling error usually shows up as marginally lower returns. An execution error of equivalent magnitude can erase the entire profit. A strategy with a clean 1.2 Sharpe and good execution will beat a 1.6 Sharpe strategy with 30% slippage erosion, every time, because the 30% erosion is taken off the top, not the bottom.

Our standing practice is to model execution costs explicitly in every backtest, and not as a flat per-trade charge. Costs are functions of instrument, time-of-day, and event proximity. The difference between a flat-cost model and a realistic one is, in our experience, often the difference between a backtest that says 'profitable' and a live result that says 'flat'.

Costs are recalibrated quarterly using our own live execution data. Backtests that pre-date a cost recalibration are not trusted at face value — they are re-run against current cost assumptions before any sizing decision is made. This is a small operational discipline that has saved us from a number of strategies we would otherwise have scaled into a deteriorating cost environment.

The default market order is, for most strategies, the wrong tool. It maximises certainty of fill at the cost of certainty of price, and for any strategy with a modest edge that trade-off is unattractive. We treat order type as a first-class design decision: limit orders with a defined patience budget, iceberg sizing for thin instruments, and time-weighted execution for larger blocks each have their place.

The trade-off is explicit. Limit orders can miss. Iceberg orders can leak information about size if the venue is poorly behaved. Time-weighted execution can pay opportunity cost during fast moves. None of these costs is hidden; we measure each one against the alternative — market orders eating the spread — and the answer is rarely 'just use market'.

Before scaling sizing on any strategy, we require that live execution matches backtested execution within a tight tolerance band across a defined live period. The check is on slippage distribution, not just average — a venue can match average slippage and still have a fat tail that destroys the strategy.

If the tolerance test fails, the strategy is paused and the execution layer is debugged before sizing is changed. A failure here is more often an execution problem than a model problem, but the discipline does not assume that. Either way the answer is to fix the underlying issue, not to scale into a strategy whose live behaviour does not match its modelled behaviour.

Execution is not where careers are made or stories are told. It is, however, where careers are quietly preserved. The traders we know who have run profitable books across decades almost universally treat execution as a serious engineering discipline — not as the boring last step after the interesting modelling work is done.

Internalising this is one of the meaningful adjustments in moving from a research mindset to a production mindset. In research, the model is the protagonist. In production, the model is one of several components that all have to work together, and execution is the component most likely to be the difference between success and failure.