Small Orders, Market Impact, and Backtest Reliability
Summary
The document examines the assumption that a small order has negligible influence on market observations and price dynamics. That assumption can make historical backtests easier because the recorded order book or other market data are left unchanged by simulated trades. The response accepts that a single small order may have limited impact, but cautions that a backtest usually represents a sequence of related orders. Their cumulative activity can affect the market and create a path different from the historical one.
It also notes that even an individual trade can alter short-horizon signals such as order book imbalance, changing the conditions for later trades. Suggested approaches include models of order flow and market impact, as well as high-frequency order book simulation; the response emphasizes that no method is perfect. It recommends assessing the scale of simulated liquidity consumption and the prices at which it occurs, then treating backtest results with corresponding uncertainty. The discussion does not quantify a universal threshold for when an order is small enough to ignore.
Key ideas
- A small order may have little immediate effect, but a series of related orders can alter market dynamics.
- An early simulated trade can change short-horizon signals used to decide later trades.
- Market impact models and order book simulations can represent some of these effects.
- Backtest confidence should reflect simulated liquidity removal and execution prices.
Tags
Full text
# Is the impact of "small" orders on market dynamics more than is commonly assumed?
# Is the impact of "small" orders on market dynamics more than is commonly assumed?
When modeling the dynamics of a market, a common assumption is that the impact of a "small" (e.g. very low percentage of daily traded volume) order on current and future observations of the market ("observations" being things one could use to determine how much of one asset you could trade for another, e.g. order book snapshots for a centralized exchange, token reserve levels for an automated market maker decentralized exchange, etc.) is almost nonexistent or nonexistent.
Among other things, this assumption then allows us to more easily use historical data for tasks such as backtesting of trading strategies as we do not have to alter the data as a result of our actions at whenever an action is taken.
What I want to know is if this is a potentially "dangerous" assumption and also what has been done to loosen this assumption since no matter how small an order is, it is still recorded in the transactions that have occurred and if those are publicly available, there is potential for situations such as other agents acting on the event of the "small" order with orders of their own that may not be so "small" in size, etc.
Would a potential remedy be to develop our models in a live environment? as in letting them forecast/trade/interact/evaluate on the actual market we're trying to model with a relatively small amount of capital in somewhat of a reinforcement learning framework? as this allows us to not need to assume how our actions might impact the marker since we see what happens in real time.
## Answer by lehalle (score 1)
https://quant.stackexchange.com/a/77990
If the assumption that a "small" order does not affect price dynamics is not false. Nevertheless, in the context of backtest, they are other aspects that cannot be neglected:
- During a back, one does not send one small order, but a series of "coherent" small orders. This series will impact price dynamics in the real life. You need a model (like a Hawkes model, cf Bacry, Emmanuel, Adrian Iuga, Matthieu Lasnier, and C-A L. "Market impacts and the life cycle of investors orders" Market Microstructure and Liquidity 1, no. 02 (2015): 1550009.) to take this into account.
- Moreover, even at the size of two consecutive small orders, if "high frequency signals" are used (like the orderbook imbalance, cf C-A L. and Eyal Neuman. "Incorporating signals into optimal trading" Finance and Stochastics 23 (2019): 275-311) the first order will probably change this signal, hence it create an "alternate future" in which the second order should be simulated, not in from of the same backtest.
They are a lot of ways to account for these effects (like C-A L., Olivier Guéant, and Julien Razafinimanana. "High-frequency simulations of an order book: a two-scale approach" Econophysics of Order-driven Markets: Proceedings of Econophys-Kolkata V (2011): 73-92. or Price Signals in Trade Execution, by Robert Almgren in 2019), and none of them is perfect. The important point is to keep enough doubts on the result you will obtain with your backtest, and to measure the "intensity" of this doubt, like the total liquidity that your simulation will remove, and at which price.Shown in full with attribution under the source's licence. Licence: CC BY-SA 4.0 (Stack Exchange)
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.