Order Book Features for Predicting Short-Horizon Midprice Changes

Summary

The document frames a feature-selection problem using order book snapshots with five bid and ask price levels, corresponding displayed quantities, and a midprice defined from the best bid and ask. The proposed target is the change in midprice across a few observation steps, illustrated by comparing the current midprice with one four rows later. The goal is to identify order book information that might help predict the magnitude or direction of that move.

The author reports trying volatility, midprice rate of change and absolute change, order quantity changes, and combinations with moving averages. In this initial exploration, using each order quantity separately performed best among the tested approaches, though no predictive results or evaluation details are supplied. The document offers a research question and a limited exploratory observation rather than a tested strategy. It does not specify the sampling interval, instrument, validation method, or whether the target is evaluated out of sample.

Key ideas

• The proposed inputs include prices and displayed quantities across five levels on each side of the book.
• The target compares the current midprice with its value several observations later.
• The author reports that separate order quantities outperformed the other tried feature groups in initial experiments.
• No sampling frequency, validation procedure, instrument, or quantitative performance evidence is given.

Tags

# Which features based on orderbook information could be relevant for price prediction?


# Which features based on orderbook information could be relevant for price prediction?












I have some orderbook data, including 5 ask prices, 5 bid prices, amount of asks and bids for every price, and midprice which is equal to (best bid + best ask)/ 2.

I would like to predict absolute change of price in a few days.

Let's say my target is:

target[i] = midprice[i+4] - midprice[i]

Which features could impact the target value?

I've tried volatility, rate of change of midprice, absolute change of midprice, order amounts change and different combinations of ROC and MA. So far, nothing has worked better than just taking all the amounts separately

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.