An execution algorithm is a scheduler with an opinion about urgency. Give it a large order and it decides how much to work now and how much to leave for later, which venues to show to, and how visible to be while doing it. Every one of those decisions is taken in service of a benchmark, which means the benchmark — not the algorithm — determines the behaviour you get.
Consider the two most common. An algorithm measured against interval VWAP is rewarded for tracking the market’s own volume profile, so it will wait when volume is thin and participate when it is heavy, and it will accept whatever the day does to the price. An algorithm measured against arrival price is rewarded for finishing before the price moves away, so it trades sooner, pays more spread, and reduces the risk of the market drifting against the order. Neither is better. They are optimising different losses, and a desk that wants the second while measuring the first will conclude, repeatedly, that its brokers are underperforming.
Implementation shortfall is the honest benchmark for most institutional flow, because it prices the two costs against each other: the market impact of trading quickly and the timing risk of trading slowly. It is also the least comfortable, since it makes the decision to delay visible as a cost rather than as prudence. Desks that adopt it tend to discover that their aversion to impact was quietly expensive.
Information leakage is the cost that sits outside all of these. An order that is predictable — constant participation rate, regular child-order sizes, the same venue sequence every time — can be inferred by other participants, and the price moves before the order completes. Randomised scheduling and size, careful venue selection and restraint about how much is shown are not sophistication for its own sake; they are what stops your own footprint becoming the adverse move you then attribute to the market.
This is why transaction cost analysis should change decisions rather than produce a monthly pack. Measure the realised cost against the benchmark you actually chose, decompose it into spread, impact and timing, and look at it by order size, urgency and venue. If the analysis never causes an algorithm to be reconfigured or a broker to be replaced, it is reporting, not analysis — and the desk is paying for the measurement without collecting the benefit.
