A Research Workflow for Neural Networks on Market Depth Data
Summary
The document offers research guidance for studying whether neural networks can predict liquidity or bid-ask spreads from market depth data. It recommends preparing the data carefully, considering zero or negative spreads, and planning how limited observations will be divided for evaluation. It also suggests first examining spread time series for stationarity and testing conventional statistical models, such as ARMA or ARIMA, before moving to neural networks.
For neural-network experiments, the suggestions include varying input features such as bid-ask volume and time, applying smoothing, and comparing network structures and outputs using both in-sample and out-of-sample performance. The guidance emphasizes stating a research hypothesis and scientific objective. It is a broad starting outline rather than a complete experimental design: it gives no data specification, implementation details, model results, or advice on leakage control and validation for time-ordered data. One answer also points toward a software learning resource, but the substantive guidance is methodological.
Key ideas
- Inspect spread data for zero and negative values before modeling.
- Plan how to divide limited observations for model development and evaluation.
- Test stationarity and conventional time-series models to understand spread behavior.
- Compare neural-network inputs and architectures, including volume and time features.
- State a research question and evaluate performance out of sample.
Tags
Full text
# Research topics - neural networks and market liquidity # Research topics - neural networks and market liquidity I am a masters student looking for some direction on using neural network on market depth data to help predict market liquidity and bid-ask spreads. Can some of the more experienced people give me some research guidance and guide me to some papers that I should be reading? I have tried Google scholar but can't find something meaningful to kick off my research. Thanks ## Answer by Greg Thatcher (score 2, accepted) https://quant.stackexchange.com/a/20851 I agree with all Robert says above, but if you already have the data, and you want to quickly create a neural network model and run the analysis, I would suggest the following: - The Heaton Site has a Wiki, links to papers, links to books, a forum, etc. that will help you get started, but you might try the PluralSight course Introduction to Machine Learning with ENCOG 3 if you want to get up to speed very quickly as this course shows you how to use Heaton's free software to create a neural network model. - Download ENCOG 3 from the Heaton site. It is a machine learning Framework that runs on a variety of languages (C#, Java, JavaScript, etc.) - In a few hours, you can get setup with code that looks something like this and have your first neural network analysis: ## Answer by Robert Szóstakowski (score 5) https://quant.stackexchange.com/a/19605 You can try using different approaches. Starting from something not that "heavy" like the NN. 0) Pre study - you need to prepare your data (how you will treat a negative spread (i.e. ASK - BID <0), what will you do if you will have 0 spread and then you will divide some value by it?), - plan your research ahead - how will you divide your limited data https://en.wikipedia.org/wiki/Cross-validation_(statistics) 1) Time series models - check if the bid-ask spread tends to be stationary (you can check it using some extracted moving period from your time series), apply statistical tests - apply ARMA, ARIMA models for periods which bid-ask spread is stationary - try to figure out what should you do when the time series stop to be stationary - try to apply mean reverting models - after doing this you will have some solid understanding of the characteristics of bid-ask spread models and it will be easier to build NN models. - this book might be useful http://www.amazon.com/Series-Analysis-James-Douglas-Hamilton/dp/0691042896 2) Neural networks - try to use different input parameters (i.e. you can smooth the data using moving average) - try to use also "bid ask volume", "time" variables - try to use different type of neural networks (different layers) and output neurons and discover which one is the best "out of sample" and "in sample" BTW what is your goal of the research? Where is the science? What do you want to discover or prove? What is your hypothesis? Answering those questions might help to plan the research in advance.
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.