- author = "Hongguang Liu",
- title = "Multi-frequency analysis for high frequency trading",
- school = "Dept. of Industrial and Systems Engineering, Hong Kong Polytechnic University",
- year = "2016",
- type = "Nov",
- address = "Hong Kong",
- keywords = "genetic algorithms, genetic programming, ANN, MLP, Wavelet MLP, NARX, WMLP, Investment analysis, Mathematical models, Portfolio management, Mathematical models, Stocks, Prices, Mathematical models",
-
URL = "
http://theses.lib.polyu.edu.hk/handle/200/8985",
-
URL = "http://library.polyu.edu.hk/record=99102195",
- broken = "http://theses.lib.polyu.edu.hk/bitstream/handle/200/8985/991021952843203411.pdf",
-
URL = "
https://theses.lib.polyu.edu.hk/bitstream/200/8985/1/991021952843203411.pdf",
- size = "xiv, 186 pages",
-
abstract = "High-Frequency Trading (HFT) in financial markets has
been making media headlines. The 2010 Flash Crash in
the US and the 2013 Everbright Securities incident in
China showed its dramatic impacts on the markets.
However, as a relatively new phenomenon, most of the
discussion on HFT is not backed by solid academic
research. At the same time, current academic research
on high-frequency trading focuses on its afterward
influences, the motivation and the trading logic behind
the HFT is rarely explored. Basically, there are two
kinds of HFT, the first kind of HFT takes advantage of
{"}time{"}, the most advanced computers are placed
right next to the exchanges to reduce the time delay of
the receiving of market data and the execution of
trading orders that aiming to capture a very small
fraction of the profit on every trade. The second kind
of HFT is conducted based on the analysis of the
historical data of the related financial time series.
This thesis focuses on the study of the second kind of
HFT. Multiple methods can be used in the design of the
second kind of HFT. In this research, multi-frequency
analysis and wavelet are combined with technical
indicators and modern machine learning tools.
Forecasting of the directions of the financial time
series is crucial in the design of such kind of HFT
systems, many economic and technical models and
indicators have been built in the past, however, most
of the past research merely analyse the data in time
domain, the frequency domain of the HFT is rarely
explored. This research focuses on the multi-frequency
predictions of the short-term movements of the
financial time series and the design of the trading
systems based on the forecast.
HFT systems based on moving averages and a simple trend following system are developed to set benchmarks for the multi-frequency related systems. An experiment on the performance of two-frequency ARIMA model is also conducted to show the prediction power of the multi-frequency analysis, as time series in different resolutions may convey different information on its characteristics, the empirical results indicated that multi-frequency could improve the forecast performance. After that, an intra-day trading system is designed based on the Genetic Programming (GP) and technical analysis, wavelet de-noise is introduced to improve the performance of the GP based system, the system with wavelet de-noise showed best performance in the empirical test. To explore the nonlinear relationship, artificial neural network (ANN) is applied in the prediction of the financial time series. Both Nonlinear Auto-regressive with eXogenous (NARX) and wavelet based Multi-layer perceptron models are used in the forecasting of the intra-day high-frequency time series, based on which, HFT systems are developed. To test the performance of the HFT systems, the China index futures is selected as the experiment asset. Based on the experiments in this thesis, the HMA trading system shows the best performance among the tested moving averages trading systems; the two-frequency ARIMA beats the traditional single frequency models; the GP systems trained using the wavelet de-noised data outperforms the GP systems trained using the original data, and the hard-threshold denoise method provides the best out-of-sample trading performance; the WMLP based trading model outperforms the NARX model in the out-of-sample trading test.",
