Scribd
Upload
116 views

Orderflow Abstract

This document proposes using input/output hidden Markov models (IOHMMs) to generate sequences of stock orders that react realistically to changing market conditions. Previous work using hidden Markov models to generate orders ignored market inputs, resulting in unrealistic trading patterns. The proposed IOHMM approach will model order generation as depending on inputs like prices and volumes, leading to more accurate order sequences under different market scenarios. Next steps involve testing an adaptive market making model within the IOHMM-generated market environment.

Uploaded by

postscript
Copyright
© Attribution Non-Commercial (BY-NC)
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PS, PDF, TXT or read online on Scribd
116 views

Orderflow Abstract

This document proposes using input/output hidden Markov models (IOHMMs) to generate sequences of stock orders that react realistically to changing market conditions. Previous work using hidden Markov models to generate orders ignored market inputs, resulting in unrealistic trading patterns. The proposed IOHMM approach will model order generation as depending on inputs like prices and volumes, leading to more accurate order sequences under different market scenarios. Next steps involve testing an adaptive market making model within the IOHMM-generated market environment.

Uploaded by

postscript
Copyright
© Attribution Non-Commercial (BY-NC)
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PS, PDF, TXT or read online on Scribd
You are on page 1/ 2

CBCL/MIT Artificial Intelligence Laboratory, September 2001 1

Input/Output Hidden Markov Models for Modeling Stock Order Flows

Adlar Jeewook Kim (with T. Poggio, CBCL)

The Problem: The project aims to develop a trainable system, which generates a sequence of orders.
For the same market conditions as the training data, the generated orders will form a similar trading
process (price and volume). The market conditions will be determined by the orders that are already
generated, and by the market maker’s actions.
Motivation: In the past few years, CBCL has developed an adaptive learning model for market-making
by using reinforcement learning (Chan and Shelton, 2001[4]). The model was tested under a simply
simulated market environment. However, testing under such simplified environment is inadequate to
argue how well the market- making model will perform under the real market. In this project, we focus
on designing a sophisticated and dynamic market environment using Input/Output Hidden Markov
Models[2] .
Previous Work: Previously, the research on the information content of the trading process has been
carried out (Easley, Kiefer, and O’Hara, 1997[6] ). In this work, they fit a model of the trade process,
which allowed to explain degrees of information content of the trading process of a particular stock.
The model was fitted by maximum likelihood using transactions data on six stocks over 60 days. They
showed that the trade process provided wealth of information as the following:

1. The large and small trades have different information content, but this varies across stocks.

2. The uninformed trades are history dependent.

3. The large buys and large sells are equally informative.

Although this is not directly related to the order flow generation process, it helps defining the structure
of our IOHMMs (particularly, input variables).
Approach: In this work, we propose to construct an IOHMM to generate a sequence of orders given
the market conditions, which will form a trading process. Each state will emit an order based on their
estimated conditional Gauissian distributions. The inputs will affect the transition probabilities distribu-
tion of the next following states, as well as the emission probabilities of the outputs (orders). The inputs
will consist of variables, which can describe the market conditions. Some of the possible inputs are
bid/ask prices, bid/ask sizes, and information content of the trading process. As a result, the IOHMM
will generate orders that react to the changes in market conditions of a particular stock used for training.
The IOHMM is defined by

• Number of hidden states m.

• Initial state distribution P0 (s0 ).

• State transition probabilities P1 (st+1 |st , market conditiont ).

• Output probabilities P0 (Ot |st , market conditiont ).

Before using IOHMMs, we preliminarily tried Hidden Markov Models [8] to model the order generating
process. Considering HMMs, we are ignoring the market conditions that might have various impacts on
the order generating process. Since we are ignoring the market conditions, the orders that are generated
by the model will form a similar trading process as the training orders. Table 1. shows volume, average
bid/ask spread, and volume weighted average price1 of trained and generated orders of three trading
days of IBM stock.

1 (ΣT P · St )/(ΣT
t=1 t t=1 St ) where Pt = last price at time t, St = traded size at time t
2

Nov. 2, 1990 Nov. 6, 1990 Nov. 9, 1990


Address Training Generated Training Generated Training Generated
Volume 356,800 465,200 294,400 440,000 576,900 515,600
Avg. Bid/Ask $0.83 $1.19 $0.43 $0.52 $1.08 $0.82
VWAP $107.55 $107.55 $107.20 $107.22 $108.74 $108.91

Table 1: HMM Training Results of IBM Using 8 Hidden States

Impact: Previously, modeling a financial market environment involved a number of unrealistic as-
sumptions such as the existence of a true price process, or differentiating the informed and uninformed
traders. And these assumptions over-simplified the model. The new approach will empirically model
the aggregated behavior of the trading crowd. This will make the adaptive learning model for market-
making possible to be tested under a more realistic market environment.
Future Work: Clearly, the next step to this project is to test the market-making model under IOHMM
market environment. Also, the results of the IOHMM model will be extensively studied by applying
various scenarios of market conditions.
Research Support: Research at CBCL is sponsored by grants from: Office of Naval Research (DARPA)
under contract No. N00014-00-1-0907, National Science Foundation (ITR) under contract No. IIS-
0085836, National Science Foundation (KDI) under contract No. DMS-9872936, and National Science
Foundation under contract No. IIS-9800032
Additional support was provided by: Central Research Institute of Electric Power Industry, Center for
e-Business (MIT), Eastman Kodak Company, DaimlerChrysler AG, Compaq, Honda R&D Co., Ltd.,
Komatsu Ltd., Merrill-Lynch, NEC Fund, Nippon Telegraph & Telephone, Siemens Corporate Research,
Inc., Toyota Motor Corporation and The Whitaker Foundation.
References:

[1] Y. Bengio. Markovian models for sequential data. Neural Computing Surveys, 2:129–162, 1999.
[2] Y. Bengio and P. Frasconi. Input/output hmms for sequence processing. IEEE Transactions on Neural
Networks, 7(5):1231–1249, 1996.
[3] Y. Bengio, V. Lauzon, and R. Ducharme. Experiments on the application of iohmms to model finan-
cial return series. IEEE Transactions on Neural Networks, 12(1):113–123, 2001.
[4] N. T. Chan and C. R. Shelton. An Electronic Market Maker. Technical Report 200-005, MIT Artificial
Intelligence Laboratory, 2001.
[5] T. Chan. Artificial Markets and Intelligent Agents. PhD thesis, Massachusetts Institute of Technology,
2001.
[6] D. Easley, N. Kiefer, and M. O’Hara. The information content of the trading process. Journal of
Empirical Finance, 4:159–186, 1997.
[7] D. Easley, N. Kiefer, and M. O’Hara. One day in the life of a very common stock. Review of Financial
Studies, 10(3):805–835, 1997.
[8] L. Rabiner. A tutorial on hidden markov models and selected applications in speech recognition. In
Proceedings of IEEE, Vol. 77, No. 2, pages 257–286, 1989.

You might also like

pFad - Phonifier reborn

Pfad - The Proxy pFad of © 2024 Garber Painting. All rights reserved.

Note: This service is not intended for secure transactions such as banking, social media, email, or purchasing. Use at your own risk. We assume no liability whatsoever for broken pages.


Alternative Proxies:

Alternative Proxy

pFad Proxy

pFad v3 Proxy

pFad v4 Proxy