Dr Jun Chen (James) completed his PhD at the
Centre for Computational Finance & Economic Agents (CCFEA),
University of Essex (2015-2019).
He was part of the Directional Change Project.
In his research, he attempted to detect regime changes using directional change based indicators.
This is a piece of interdisciplinary research in the field of computational finance,
covering finance, machine learning and data science.
This work applies machine learning to financial market monitoring and algorithmic trading.
The research is based on “directional change”,
a new data-driven approach in financial data analysis.
James passed his viva on 7th May 2019.
He was examined by Frank McGroarty (External Examiner, University of Southampton), Bart De Keijzer (Internal Examiner) and Abdel Salhi (Chair).
His PhD qualification was confirmed on 30th October 2019.
Studying Regime Change using Directional Change
What are regime changes?
Financial markets reflect what is the collective trading behaviour of traders. Such behaviour is often affected by financial crisis or political events. The term regime change is used to describe such significant change of collective behaviour. This thesis studies how regime changes can be measured and detected in financial markets.
What have we done?
The traditional ways to detect regime changes are based on analysis of the statistical properties of time series. For example, researchers may have used significant changes in means, volatilities, autocorrelations and cross-covariances of asset returns to conclude regime changes.
In this thesis, we study regime change detection using indicators developed in Directional Change (DC). DC is an alternative way to sample financial data. Unlike time series, which samples transaction prices at regular time intervals, DC samples prices at peaks and troughs of the market. This method of sampling is known as zigzag in technical analysis.
What have we achieved?
Regime changes can be detected under the DC framework:
We propose a new method to detect regime changes under the DC framework. DC data is fed into a Hidden Markov Model (HMM), a machine learning model, which aims to discover the hidden state of the market. To evaluate our method, we apply it to the Forex market over a time period of uncertainty, namely the Brexit referendum period. The timing of regime changes detected by this method is consistent with the political developments taking place at the time. While regime changes detected by DC and time series agree with each other most of the time, some regime changes found under DC were not found under time series. That means our DC approach complemented the time series approach by the provision of supporting and additional information.
Normal and abnormal regimes are clearly separable in the DC-indicators space:
With the method developed, we then went on to detect normal and abnormal market regimes (which represent regimes before and after significant events took place) in other assets. Through observation of regimes detected in ten different markets at different time using different thresholds, we discovered that normal and abnormal regimes are clearly separable from each other in the DC indicator space. This allowed us to generalise and characterise what are the features of normal and abnormal market regimes using DC indicators.
With the regimes characterised, we can track the market for regime change:
Finally, we showed that the regime characteristics established above can be used for regime tracking.
As a proof of concept, we showed that, based on the market data observed so far,
one can use a simple Bayes model to compute the probability of the current market being in the normal or abnormal regime.
Preliminary results suggested that the proposed method managed to detect regime change signals accurately and promptly.
To summarise: this thesis pioneers a new method for regime change detection under the DC framework.
It showed that normal and abnormal regimes can be characterised using DC indicators.
Once such characteristics are clearly established, they could be used for market tracking.
Thus, this research lays the foundation for building financial early warning systems.