Live Execution Cost Monitoring

A backtest assumed 8 basis points of round-trip cost per trade, based on spreads measured a year earlier. The strategy launches, and for three months live costs track close to that assumption — nobody is watching closely enough to notice they're drifting to 11 bps. By month six they're at 15 bps, because the strategy's own trading, now running at real size, has started moving the names it trades. Nobody built a monitor to catch this, because the backtest's cost number was treated as a settled fact rather than a forecast needing daily checking.

What to compare, and against what

The building block is implementation shortfall: the difference between the price when the decision to trade was made (the "arrival price") and the price actually achieved, including fees. Live monitoring tracks this number every single day, broken into pieces:

$$\text{shortfall} = \underbrace{(\text{arrival} - \text{decision})}_{\text{delay cost}} + \underbrace{(\text{execution} - \text{arrival})}_{\text{market impact}} + \underbrace{\text{fees}}_{\text{explicit cost}}.$$

In words: shortfall splits into cost from waiting too long to start trading, cost from the market moving against you while you traded, and the fees and spread you paid outright. Each piece has a different fix — delay cost points at scheduling, impact points at order sizing and venue choice, fees point at broker or venue selection — so lumping them into one number hides which lever to pull.

Worked example: catching creeping impact

A strategy's backtest assumed a flat 8 bps round-trip cost. Live daily shortfall is logged for 60 days. Weeks 1–4 average 8.6 bps — within tolerance. Weeks 5–8 average 11.2 bps. A control-chart rule (flag any 10-day rolling average more than 2 standard deviations above assumption, with historical day-to-day SD of 1.8 bps) trips at week 6's $8+2(1.8)=11.6$ bps threshold. Decomposing shows delay cost flat at 1.5 bps, fees flat at 2 bps, but market impact climbing from 4.5 to 8.3 bps — participation rate has grown from 3% to 7% of daily volume as assets grew, and impact scales with the square root of participation. The fix is a capacity cap, not a broker change.

Worked example: a fee change hiding in the total

A different strategy's total shortfall looks stable at 6 bps for months, clearing the same check with no flag. A quarterly deep-dive finds market impact actually fell from 4 to 2 bps — the signal decayed, so it traded smaller, less-informed positions — while fees rose from 2 to 4 bps after a venue-routing change. The total stayed flat by coincidence, masking two real, offsetting problems.

The control-limit logic above mirrors that plot's confidence band: a rolling average has its own sampling noise, so the alert threshold sits outside the normal range of day-to-day cost variation, not at the point estimate, or you get false alarms every week.

What this means in practice

Log arrival price, execution price, and fees for every order, decompose shortfall daily, and alert on each component against a control limit tied to its own historical volatility. Feed results back into Capacity-Constrained Backtesting so cost assumptions update as the strategy's footprint changes.