Page 121 - Undergraduate Research Opportunities Program Proceedings 2017-18

School of Business and Management
Department of Information Systems, Business Statistics & Operations Management

Department of Information Systems, Business Statistics & Operations Management

Family Consent and Organ Donation

Supervisor: HUI Kai Lung / ISOM
Student: ZHAI Qinwen / MAEC Course: UROP1000, Summer

Objectives: To investigate what effects family consent has on organ donation rate across different countries through
certain years.Data sources: The organ donation rate data are from International Registry in Organ Donation and
Transplantation (IRODaT). Law information is collected from official website or documents from different sources listed
in the appendix file 2. Conclusion: Need of family consent when the deceased’s wish is unknown legally shows negative
effects on donation rate from regression results. However, due to constraints of our model, there may be other omitted
factors such as culture, religions and infrastructure in healthcare that can influence donation rate, thus reduce the
extent family consent can account for the differences in donation rate across countries.

Deep Learning for Stock Prediction

Supervisor: KWOK James Sai Ho / ISOM
Student: LAU Yiu Lung Peter / ECOF Course: UROP1100, Fall
UROP2100, Spring
LUI Leung Yau / QFIN UROP1100, Spring

In this report, the applications of Discrete Wavelet Transform (DWT) and Stacked Autoencoders (SAEs) on pre-
preprocessing financial data are studied. Preliminary testings of DWT and SAEs on S&P 500 data and MSCI China data are
conducted. Details on the data collection, model architecture, optimization process and preliminary outputs are
presented. We found that DWT is useful for denoising financial time series by simply attenuating all detailed coefficients
can reduce the variations. However, it cannot avoid disposing of the huge ups and downs of the financial time series. For
SAE, we found that it is a good feature extraction tool in a sense that it can extract the most important features
representing the variation in the original dataset (i.e. the deep features). Future research directions are also proposed.

Supervisor: KWOK James Sai Ho / ISOM
Student: LIU Weiyang / QFIN Course: UROP1100, Summer

This paper explores a deep learning approach implementing Long Short-Term Memory (LSTM) network to predict a next-
minute market direction, based on 1-minute historical data of prices and volume of stocks on the Hong Kong Stock
Exchange (HKEX). We evaluate some technical indicators derived from price and volume and find that they do not
contribute much to this binary classification problem. Discrete Wavelet Transform, a common denoising method, brings
little improvement either. We also conduct tests on tuning different hyperparameters but have not discovered any
significant ones yet. A two-layer LSTM network has an acceptable performance of 76.02% directional accuracy on stocks
traded frequently but fails to predict the Hang Seng Index. Its application to forecasting lower-frequency stocks requires
more investigation.

111