Pairs trading with long-short term memory

Loading...
Thumbnail Image

URL

Journal Title

Journal ISSN

Volume Title

School of Business | Master's thesis

Date

2019

Major/Subject

Mcode

Degree programme

Finance

Language

en

Pages

30 + 13

Series

Abstract

The cross-disciplinary paper explores the applicability of different neural network architectures in financial time series prediction, bridging the still rather wide gap between machine learning and finance. Three network architectures of increasing complexity - feedforward, recurrent and long-short term memory - are presented while justifying why the additional model complexity is necessary for market data. A long-short term memory model is then chosen according to the requirements of the data characteristics. The specific input variables of the model are then chosen based on existing pairs trading literature and the model is implemented and iteratively tuned to fit a previously explored pair, predicting it's spread. The architecture is then validated on a larger sample picked via a cointegration metric before finally eliminating look-ahead bias by testing on a new period on a new subset of stocks. The computationally inexpensive model captures effects associated with pairs trading observable throughout the sample and forms a portfolio with large excess returns significant through early 2019, suggesting that excess returns can still be generated within comoving stocks using advanced nonlinear methods.

Description

Thesis advisor

Suominen, Matti

Keywords

neural networks, pairs trading, machine learning, long-short term memory

Other note

Citation