SignalTrace

Transaction Costs and Slippage

John Montgomery6 min read

Transaction Costs and Slippage

Transaction costs and slippage are the trading frictions — broker commissions, the bid–ask spread, market impact, and the gap between the decision (decision/arrival) price and the achieved fill price — that stand between a model’s gross signal and its net tradeable return. They are the single most common reason a positive Markov-model backtest fails to translate into profit. The components are not interchangeable: commissions are a small, near-fixed per-trade fee; the half-spread is the cost of demanding immediate liquidity; market impact is the price move a trade itself causes by consuming finite order-book depth; and slippage is the residual gap between the price the model assumed and the price actually obtained, which widens with order size, volatility, and latency. Crucially, only commissions behave like the flat “X basis points per trade” assumption used in most backtests — market impact does not.

The structural problem for Markov-family strategies is turnover. Because market impact scales with order size relative to liquidity, total cost drag is roughly turnover × cost-per-unit-traded, so a strategy that rebalances often pays the friction many times over. The in-vault evidence shows this directly: Shu Yu and Mulvey 2024 report HMM-guided turnover of 141–290% per year, and after a 10 bp one-way cost the HMM signal’s net return fell below buy-and-hold, while the lower-turnover Statistical Jump Model (turnover as low as 44%) stayed ahead — the cost gap, not the regime logic, drove the ranking. Bulla et al. 2010 only preserved a positive post-cost edge by deliberately engineering turnover down, and even then the out-of-sample excess return over buy-and-hold was a modest 18.5–201.6 bp/yr. Whether a paper includes realistic costs and slippage is therefore a primary evidence-grading criterion for every paper and backtest-result note in this vault.

How costs are modelled matters as much as whether they are included. Almgren Thum Hauptmann Li 2005, fitting a market-impact model to Citigroup US equity-desk data, decompose impact into a permanent component (a lasting, roughly linear price shift carrying the trade’s information) and a temporary component (a concave, mean-reverting cost of demanding liquidity); they reject the pure square-root law for temporary impact in favour of a 3/5 power law over their order-size range, while the broader Square-Root Law of Market Impact remains the standard reduced-form approximation. The common theme is that impact rises with the trade’s size relative to average daily volume and with volatility — so a flat basis-point fee is structurally wrong for any non-trivial position size and silently understates costs for large or fast trades. Abbade and Reali Costa 2026 demonstrate the practical consequence for reinforcement-learning / MDP trading: in open-source RL backtest environments a flat 10 bp fee lets agents learn pathological high-frequency trading; switching to an Almgren–Chriss impact model cut one agent’s daily costs from 8k as turnover collapsed from 19% to 1%, and dropped a tuned PPO agent from 34% to 25% out-of-sample return. The cost model is part of the environment, so it changes which policy is learned at all — not merely the reported P&L of a fixed policy.

How large the drag is in practice is genuinely contested, and the two anchor studies bracket the range. Frazzini Israel and Moskowitz 2018, using over $1 trillion of live institutional trades from AQR Capital Management across 21 markets, measure mean market impact of only ~9–11 bp per trade — an order of magnitude smaller than quote-based estimates — and conclude that size, value and momentum survive costs at large fund sizes (short-term reversal does not). But those are the costs of a sophisticated desk that supplies rather than demands liquidity. Patton and Weller 2017, measuring the gap between paper factor returns and what mutual funds actually deliver, find annual implementation costs of 2.2–8.5% and conclude that for typical funds the return to momentum — the canonical high-turnover anomaly — is statistically indistinguishable from zero after costs. The lesson for Markov strategies is sober: a regime-switching signal with momentum-like turnover may be perfectly real on paper and still net to nothing for any trader without elite execution. Realistic cost and slippage modelling is therefore a load-bearing assumption, not a footnote — and a backtest that omits it, uses a flat fee, or assumes mid-price fills cannot be graded above weak on this vault’s profitability scale.

Transaction Costs and Slippage [contradicts] Markov Chain Trading Model Transaction Costs and Slippage [opposes] Hidden Markov Model Regime Detection Transaction Costs and Slippage [relates] Out-of-Sample Backtesting Square-Root Law of Market Impact [defines] Transaction Costs and Slippage Almgren Thum Hauptmann Li 2005 [supports] Square-Root Law of Market Impact Frazzini Israel and Moskowitz 2018 [contradicts] Patton and Weller 2017

Connections

Sources

Graph View

Backlinks