Statistical and Hawkes Models for Anticipating Order Flow
Summary
The document surveys ways to estimate the direction and size of upcoming order flow from intraday market data. It distinguishes behavioral models, which simulate trader actions, from statistical approaches that use tick-level order book variables and other order data as predictors. The choice of target matters because order flow can refer to several distinct quantities derived from trades and the book.
The answers name Poisson processes as a simple starting point, while cautioning that this approach is naive. They also point to Hawkes processes for modeling clustering in trades and to autoregressive conditional duration models for event timing. Agent-based simulations may rely on unrealistic assumptions, and any forecast depends on the specific order-flow measure and available depth and order data. The document offers model families and research directions rather than a tested forecasting setup or evidence of predictive performance.
Key ideas
- Order flow can refer to several different variables, so define the forecast target first.
- Statistical models can use tick-level order book measurements and order data as predictors.
- Poisson processes offer a basic but limited starting point for modeling order flow.
- Hawkes processes can represent clustering in trade activity.
- Duration models such as ACD provide another approach to modeling event timing.
Tags
Full text
# Order anticipation # Order anticipation Is there a way to anticipate order flow on a security. For simplicity's sake i'm referring to a security that is traded on one exchange and has a single order book, by anticipating order flow i mean is there any way to evaluate the size and direction (buy / sell) of the next order placed in real time? ## Answer by develarist (score 2) https://quant.stackexchange.com/a/53439 If by order flow you mean high-frequency changes in prices, returns, volume, and other variables based on intraday data at various levels of market depth, yes there are various approaches that have been developed, typically falling under two categories: behavioral approaches, which try to model order flow by simulating trader behavior, often with agent-based models that entail unrealistic assumptions, and statistical approaches which rely on quantitative measures that often take into account averaged tick-by-tick distributions of the limit order book variables by using order data besides the target variable for features. The Poisson process is a common but naive starting point for modeling order flow. Order flow can mean any one of a number of variables taken or derived from the order book though, so it will all depend on which one you want to anticipate. Quantitative finance publishes research on order book models. ## Answer by ltrd (score 0) https://quant.stackexchange.com/a/53445 Welcome BlackStar, It is really interesting topic to modelling orderbook dynamics and trade dynamics. There are a lot of works related to the family of processes called Hawkes processes. Those models connected with trades try to find information using clustering behavior of the trades. Also duration processes such as ACD processes should be interesting for you. You should also consider retrieving information from orders and liquidity in orderbook but here I will leave you with this topic alone. Interesting applcation of the Hawkes process was used in paper "Modelling trades-through by Hawkes process" where author try to model market orders that went through (I do not know if this is appropriate expression) first line in orderbook.
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.