[Insight 06] ·
Notes from the desk
Order lifecycle accounting
An order is most often evaluated as a single event with a binary outcome. It either filled at the requested price or it did not. When measurement is more careful, that binary is replaced with a single number — slippage against arrival mid, or implementation shortfall against decision price — and the order is closed in the books. The number is reported, the desk moves on, and the stage at which the cost was actually incurred is left unattributed.
The fuller framing is that an order is a sequence of distinct stages, and each stage carries its own cost. Some of those costs are explicit and appear on a fee schedule. Others are implicit, contingent, and visible only when the comparable decision was made differently on a similar order moments earlier. A summary metric that collapses all of these into one figure is convenient for reporting and almost useless for improving execution. The corrective is to account for each stage separately, in the same currency, and on every order — not just the ones that look unusual after the fact.
Signal generation is the first stage and the easiest to mis-attribute. The cost here is opportunity: the elapsed time between the market state that justified the order and the moment a decision crystallised. In quiet conditions the cost is negligible. In fast conditions the same elapsed milliseconds can move the reference price past the level the signal was meant to capture, so that even a perfect downstream implementation acts on a stale view. Desks that benchmark only against arrival mid will not see this cost at all, because arrival mid is measured after the decision has already been made.
Route selection is the second stage and the one most often left implicit. Choosing a venue is choosing a distribution of expected fill quality, queue position, and adverse selection. The conventional treatment evaluates routes on realised fills against the chosen venue’s own touch, which answers the wrong question. The relevant counterfactual is what the same order would have done on the alternatives that were not chosen. Without an estimate of that counterfactual — even a coarse one — the cost of the routing decision is invisible, and the route stops being a managed parameter.
The submission stage carries a cost that is structurally different from the others. Between the moment an order is generated and the moment it is acknowledged by the venue, the market continues to move. That drift is not slippage in the conventional sense — it occurs before the order has any presence in a book — but it directly determines whether the order arrives in front of, alongside, or behind the price level it was designed for. On venues with strict price-time priority, a few milliseconds of drift can be the difference between a fill at the touch and a fill several ticks worse, or no fill at all. The cost is real and asymmetric: drift in the desk’s favour collapses to a fill at the original price; drift against it shows up as either a worse fill or an unfilled order whose opportunity cost is then attributed to signal quality rather than to submission latency.
At the matching stage the explicit costs finally appear. Maker rebates, taker fees, and the spread itself are visible and easy to attribute. The implicit cost at this stage is queue dynamics: an order resting in a book has a position that can be invalidated by a single sufficiently large counterparty arrival, and the realised fill rate of a passive order depends as much on queue health as on the price chosen. Treating queue position as a constant rather than a parameter that varies across venues and across the day misses one of the larger sources of variance in fill quality.
Settlement is the stage where attribution is most often skipped, because the costs feel external to the trading decision. Network gas on a decentralised venue, custodial transfer cost on a centralised one, clearing and netting timing on either — these are usually accounted for at month-end, on a venue total, rather than per order. That treatment is defensible for back-office reporting and indefensible for execution analysis, because settlement cost is not invariant: it correlates with venue choice, with order size, and with network conditions, all of which are decisions made earlier in the lifecycle. Pulling settlement cost back to the order level closes the loop.
The practical implication is that an aggregate fill-quality number is a starting point, not a conclusion. A desk that improves its overall slippage by twenty per cent without knowing which stage the improvement came from has no basis for the next decision: whether to keep investing in signal compression, route diversification, submission latency, queue management, or settlement optimisation. Per-stage attribution is the prerequisite for any of those choices being deliberate rather than incidental. It is also the prerequisite for the harder conversation about which stages are worth optimising further at all — because beyond a point, every stage exhibits diminishing returns, and the marginal pound is better spent on the stage that has not yet been measured.