At 9:42 a portfolio manager decides to sell.

The model has done its job. The position is too large for the new forecast; the expected return no longer compensates for the risk; the target portfolio is clear. On paper, the trade is already finished. Sell \(Q\) shares at the decision price \(S_0\), update the book, move on.

The market does not honor that fiction.

The desk now owns the part the backtest left blank: how much to release, when to route it, how aggressively to cross the spread, how much residual inventory to carry, and how much information to reveal. The realized portfolio will not be marked at the decision price. It will be marked at the prices created by that path.

Perold’s implementation shortfall is the clean name for this gap between the paper portfolio and the implemented portfolio.1 The phrase sounds accounting-like, but it is really a warning: alpha is not owned until it is executed.

Desk rule: the decision price is a coordinate. Execution is the curve drawn through market liquidity to reach the target portfolio.

The Arrival Price Is a Witness

For a buy order, a stripped-down shortfall calculation is

\[\text{shortfall} = \frac{1}{Q}\sum_{i=1}^N q_i S_i - S_0,\]

where \(q_i\) is the child order filled at step \(i\), \(\sum_i q_i = Q\), and \(S_i\) is the realized execution price. For a sell order the sign changes, but the benchmark role of \(S_0\) is the same. It is the arrival price: the price at which the paper decision became an execution problem.

This number is richer than a commission line item. It mixes spread, temporary impact, permanent impact, delay, adverse selection, incomplete fills, and the opportunity cost of whatever remains undone. It also prevents a convenient mistake. A strategy with beautiful paper alpha but terrible implementation is not a high-Sharpe strategy with costs. It is a different strategy.

Bertsimas and Lo put this into a dynamic optimization frame: choose a sequence of trades over a finite horizon to minimize execution cost under a price-impact model.2 Almgren and Chriss added the mean-variance geometry that many execution discussions still orbit: some schedules minimize expected cost; some minimize uncertainty; the useful object is the frontier between them.3

Footprint, Weather, Nerves

There are three forces in the simple version of the problem.

The first is footprint. A large child order consumes liquidity and pushes the execution price away from the arrival price. In the clean quadratic model, if \(x_i\) is remaining inventory after step \(i\), then temporary impact behaves like

\[\sum_{i=1}^N (x_{i-1} - x_i)^2.\]

This term likes patience. It says: split the order, avoid concentration, do not slam the book unless you must.

The second is weather. While inventory remains, prices move. A stylized inventory-risk term is

\[\sigma^2 \sum_{i=1}^N x_i^2 \Delta t.\]

This term likes urgency. High volatility, long horizons, and institutional drawdown limits all make residual inventory expensive.

The third is patience with a sign. If the asset has favorable expected drift while you are selling, waiting can help. If the drift is adverse, waiting is a tax. The same order size can deserve a different path when the trade is liquidity-motivated, risk-motivated, alpha-motivated, or forced by a margin constraint.

Those forces are why “minimize transaction costs” is too vague. A schedule that minimizes expected footprint may maximize exposure to price movement. A schedule that minimizes risk may pay too much impact. Execution is the art of choosing which regret you prefer.

The Book Remembers You

The simple model treats each child order as if it pays a local cost and then walks away. Real books remember.

Obizhaeva and Wang make this memory explicit by modeling supply and demand as dynamic objects. The key state variable is resilience: how quickly liquidity recovers after a trade consumes it.4 Two assets with the same spread and displayed depth can require different execution if one book refills quickly and the other stays wounded.

That single word, resilience, changes the mental model. Market impact is not just a function from size to cost. It is a system response. A child order changes the state in which the next child order arrives.

Recent empirical work keeps pushing in the same direction. Maitrier, Loeper, Kanazawa, and Bouchaud use Tokyo Stock Exchange data with trader identifiers and argue that the square-root law of impact has microscopic roots: impact appears already at the child-order level after the market digests the trade, and synthetic metaorders reconstructed by scrambling trader identity display similar behavior.5 The lesson is not that one universal formula solves execution. The lesson is that impact has mechanical structure, and the structure has to be estimated rather than hand-waved.

Round Trips Should Bleed

An execution model should fail loudly if it allows profitable round trips.

Gatheral’s no-dynamic-arbitrage principle says that expected trading costs should not become negative just because a trader buys, sells, and harvests their own market impact.6 Gatheral and Schied show why transient-impact models need regularity conditions: plausible-looking decay functions can create price manipulation or ill-behaved optimal strategies.7

This is where machine learning needs humility. A policy network can predict fills, choose venues, and adapt to state. But if the learned simulator rewards churn, or if its impact model makes round trips profitable, the model has learned an accounting bug in the market universe it was given.

The audit question is simple:

\[\text{expected cost of a round trip} \ge 0.\]

If that inequality fails, do not tune the policy. Fix the world model.

A Small Execution Desk

The lab below solves a discrete liquidation problem. It is deliberately small: one asset, uniform time buckets, quadratic temporary impact, inventory risk, and an optional drift term while the residual position is still held.

The optimizer chooses an inventory path \(x_0,\ldots,x_N\) with \(x_0=1\) and \(x_N=0\). The chart shows that path against VWAP and plots the cost-risk frontier produced by changing risk aversion. Costs are in basis points of traded notional; risk is the standard deviation of implementation shortfall caused by holding residual inventory during the horizon.

Under the hood, the code solves the first-order conditions of a quadratic program. For the interior inventory points, the equation is tridiagonal:

\[-\eta x_{k-1} + (2\eta+\lambda\sigma^2\Delta t)x_k - \eta x_{k+1} = \frac{\mu\Delta t}{2}.\]

The left side says that the path should be smooth when temporary impact \(\eta\) is expensive, but should shrink inventory when risk aversion \(\lambda\) and volatility \(\sigma\) are expensive. The right side is drift \(\mu\): positive drift while selling makes waiting less costly; negative drift makes waiting hurt. The JavaScript implementation solves this tridiagonal system directly, then recomputes the frontier by sweeping \(\lambda\).

Optimal inventory VWAP inventory Current frontier point Cost-risk frontier

Deterministic toy model. Costs are expressed in basis points of traded notional; risk is the standard deviation of implementation shortfall from holding residual inventory during execution.

Try the lab as if you were at the desk:

  • Raise daily volatility. Inventory risk gets louder, so the optimal path moves forward.
  • Raise impact slope. Concentrated child orders become expensive, so the path flattens.
  • Push drift negative. Waiting now has adverse expected carry, so the model accelerates.
  • Push drift positive. Waiting becomes less painful, and the optimizer is more willing to let footprint fall.

VWAP is useful as a benchmark because it is legible. It is not sacred. In the frontier view it is simply another point, often reasonable, sometimes dominated, and always conditional on the market state the model assumes.

The Blotter I Would Want

If I were reviewing a real execution system, I would want the post-trade file to look like an experiment log, not a screenshot of a fill summary.

Record the arrival price, decision time, release time, completion time, and benchmark. Record every child order: size, venue, limit price, fill, cancel, queue outcome, and timestamp. Record participation rate by interval relative to realized volume. Separate spread paid, temporary impact, permanent impact, delay, and residual opportunity cost. Compare the pre-trade impact forecast with realized shortfall. Measure post-trade reversion at multiple horizons. Re-run the same impact model on a buy-sell round trip and check that churn loses money.

The point is not bureaucracy. It is identifiability. Without this log, the desk cannot tell whether a bad fill came from bad alpha, bad urgency, bad venue selection, bad impact calibration, slow book resilience, or simple bad luck.

The Desk I Still Want

The execution research program I trust would combine learning with constraints. It would let policies adapt to order-book state, but it would not let them violate inventory conservation or no-dynamic-arbitrage. It would report uncertainty about impact decay, not only a recommended slice. It would separate alpha urgency from liquidity urgency. It would stress every recommendation under wider spreads, slower resilience, higher volatility, lower hidden liquidity, and information leakage.

The dream is not an algorithm that always trades fast or always hides. The dream is a system that knows what kind of market it is currently inside.

That is why a trade is a trajectory. Prediction decides where the portfolio should go. Execution decides the path, and the path writes itself into returns.

Paper Trail

  1. Andre F. Perold, “The Implementation Shortfall: Paper vs. Reality,” Journal of Portfolio Management, 1988. Harvard Business School. ↩

  2. Dimitris Bertsimas and Andrew W. Lo, “Optimal Control of Execution Costs,” Journal of Financial Markets, 1998. MIT abstract. ↩

  3. Robert Almgren and Neil Chriss, “Optimal Execution of Portfolio Transactions,” Journal of Risk, 2000. Journal of Risk. ↩

  4. Anna Obizhaeva and Jiang Wang, “Optimal Trading Strategy and Supply/Demand Dynamics,” Journal of Financial Markets, 2013. MIT PDF. ↩

  5. Guillaume Maitrier, Gregoire Loeper, Kiyoshi Kanazawa, and Jean-Philippe Bouchaud, “The double square-root law: Evidence for the mechanical origin of market impact using Tokyo Stock Exchange data,” 2025. arXiv. ↩

  6. Jim Gatheral, “No-Dynamic-Arbitrage and Market Impact,” Quantitative Finance, 2010. SSRN. ↩

  7. Jim Gatheral and Alexander Schied, “Dynamical Models of Market Impact and Algorithms for Order Execution,” in Handbook on Systemic Risk, 2013. SSRN. ↩