Slippage

How it works

Start with the live measurement, because it is the only version of this quantity that is a fact rather than a choice.

Execution quality in US equities is disclosed against the quote in force when the order arrived. The SEC's order execution disclosure rule defines average effective spread as the share-weighted average of, for buy orders, double the difference between the execution price and the midpoint of the national best bid and national best offer at the time of order receipt, and the mirror image for sells; realised spread is the same arithmetic against the midpoint at a specified interval after execution. Those two numbers separate the cost paid on arrival from what the price did next — and both of them require knowing the NBBO at a timestamp, per order.

A backtest has none of that. No order arrived anywhere, so there is no quote at time of receipt, no queue position, and no counterparty. What the engine has is a row of price data and a rule you gave it for degrading the fill. That rule is the slippage model, and it comes in four shapes:

Fixed. A constant number of ticks, points or cents added to a buy and subtracted from a sell. Simple, reproducible, and indifferent to the size of the order and the state of the market.

Percentage. A fraction of the fill price. Scales with the price level, still indifferent to liquidity. Composer backtests with a default assumption of one basis point on daily adjusted closing prices, which is the whole model.

Spread-based. Half the quoted spread charged on each side, which is the closest of the four to the regulatory definition above — and it needs a spread. bt models this only if you hand the backtest a bid/offer frame alongside the prices; without one, slippage is whatever cost model you wrote.

Volume-share or market-impact. Cap the fill at some share of the bar's volume, and optionally make the price concession a function of the size taken. Backtrader has volume fillers that cap a fill at a proportion of the bar's volume, plus percentage or fixed slippage with separate switches for whether opening prices and limit orders slip at all.

And a fifth shape worth naming because it looks like the others: nothing. QSTrader accepts slippage_model and market_impact_model arguments on its simulated broker and then overwrites both with None, each marked in a comment as still to be implemented. A parameter that is accepted and discarded is worse than an absent one, because the code reads as though it was configured.

The data most tests run on cannot support any of the four

A daily or minute bar carries an open, a high, a low, a close and a volume. It does not carry the quote, the spread, the depth, the sequence of trades inside it, or how much of that volume was available at the moment your hypothetical order arrived. Every model above is therefore either indifferent to those facts or inventing them.

The facts exist, at a price and a volume. The NYSE Daily TAQ product is specified as every quote and every trade with exchange timestamps — which is what a spread-based or queue-aware fill assumption would need, and what tick data means when a vendor is using the expensive sense of the phrase. Running a bar-level test and then reasoning about spread capture is reasoning about a number that is not in the file. A vectorised simulator makes this concrete: VectorBT PRO still fills at bar prices, so intrabar sequencing and queue position are outside what the run can express, whatever slippage figure is passed in.

Why it matters here

Slippage is the input in a backtest that most directly scales with turnover, which makes it the input that most changes the ranking between two rules. Across backtesting frameworks the defaults differ, the units differ — a percentage in one library, absolute ticks in another — and the setting usually lives in the API reference rather than anywhere a reader will meet it. A figure copied from one framework into another without checking the units is not a conservative assumption; it is a different assumption of unknown size.

So the useful form of the question is not "what is the slippage" but "what did this run assume, in what units, and could the data underneath it have supported anything better". A result is a function of that assumption, and reporting the result without it is reporting half of the arithmetic.