Advices to students
Advices to undergradutes
Take minutes to guide progress (your supervisor may have other template, follow theirs if so)
Advices to project workers
On research
On project
Billy Yu is an Associate Professor in the Faculty of Business at Macao Polytechnic University. He began his career as an e-business consultant serving multinational corporations before transitioning to academia, where he has led numerous government-funded and commercial research projects. At the university, Billy lectures courses aligned with the SAS® Certified Predictive Modeler credential, globally recognized qualification in analytical CRM and data analysis. Many of his students have achieved this certification under his guidance. His teaching is closely tied to his research interests in technology applications in economics and business.
Since January 2024, he has served as an editor for PLOS ONE; and occasionally a guest editor for Frontiers in Psychology. He also serves on the board of the Information Systems Audit and Control Association (ISACA) Macao Chapter, a role he has held for over 10 years.
* Browse SAS for our recognition the Academic Specialization in Business Intelligence
余达纬在澳门理工学院受聘为副教授。于香港大学获得博士学位。曾担任电子商贸顾问，其后转往学术界发展。凭顾问期间累积的经验，他曾授聘於多个由政府或商界资助的研究项目；范围主要集中在经济及管理方面。在商界，他也讲授行政课程及参与主题演讲。余采用的研究方法有定性的、也有定量的。他在问卷设计和数据分析等都有丰富经验，他的科学分析有助商业实务中作出清晰的决择。学术上的历练使他善于掌握分析工具：SPSS、LISREL和EViews；他其中一数据挖掘课更被SAS®*认可。
Doctor of Philosophy (Ph.D.), the University of Hong Kong.
Bachelor of Engineering (B.Eng.), the University of Hong Kong.
My interests are many. Behavioral economics and monetary economics are my most recent pursuits. Yet, I have strong favors in empirical investigation. I seek truths by means of statistics or even analytics, and do not limit myself to any quantitative methods. This is my school of thoughts – quantitative management. It is an inheritance from Peter Drucker:
"If you can't measure it, you can't improve it."
I even instill this idea to my students and they are strong in data mining. Well, that is where e-commerce comes into play after the millennium.The general public use paper-based receipts in the Macao retail industry; and customer always collect stickers for various kinds of retail promotions and gifts. Losses, handling and frauds in operations are major problems for residential applications. This project proposes an electronic means for that, and with better trust and reliability. Blockchain technology is used to achieve the goals for E-receipt and E-sticker. Blockchain technology provides a decentralized and open mechanism to support next-generation Internet technologies, including anonymous online payments, remittances and other digital asset transactions. By using industrial standard development tools, the students implemented a prototype for the identified operations involved. They proved their solution concepts by programming in blockchain. Their ideas not only improve security, reduce paper use, but also automate the process of receipt-to-stickers, and make possible the transfers and the tasks much more convenient.
Supply Chain Management (SCM as Michael Porter described) refers to the activities and processes of planning, coordinating, operating, controlling and optimizing the whole supply chain system, including all the stakeholders like suppliers, manufacturers, distributors, and consumers. In it, a critical challenge is the bullwhip effect, in which demand order variances are amplified as they move upstream. By identifying the major characteristics and solution paths in the literature, a blockchain framework is proposed to explicitly solve the problems. By using industrial standard development tool, the students programmed a prototype to illustrate the solutions they proposed. It proved the concepts of solving the SCM problems by using blockchain technology with its unique features. Not only it increases efficiency just as other electronic means, it also benefits the whole supply chain by improving information sharing, the inventory policy and the demand forecast and ultimately reducing costs.
Reviews are important for all consumers nowadays. This user-generated content has got serious attentions for all marketing businesses, from simple brand management, customer acquisition or even risk management. Manual process is not scalable for the massive textual data. This project deployed multiple deep learning language models to assist their analysis on over 400,000 reviews messages in TripAdvisor. Simple as basic RNN (recurrent neural network) like LSTM (long short-term memory) and sophisticated as stat-of-the-art BERT (Bidirectional Encoder Representations from Transformers) were used with different preprocessing treatments. With those NLP techniques, they could accurately estimate the sentiment of a visit from the user comment or sharing. More, the study identified important aspects for both positive and negative reviews. It provided insights for marketing practitioners to adjust their services from the massive user reviews without sweat.
Information overloading is an issue executives are faced with in today's business settings. Tons of news are to be digested for prudent decision making every day. While time is a constraint, the fullness of information in news articles should not be discounted. Summarization should be able to help. This study deployed multiple methods for the job: clustering, transformer, BERT (Bidirectional Encoder Representations from Transformers) and GPT (Generative Pre-Training Transformer). To assess the models, the author also investigate the loss functions, with consideration on feasibility and practicality. They did a straight-forward and relatively easy solution with extractive summarization by BERT. By tailor-making a pre-trained model with clustering technique for Chinese tokenization and embedding, extractive summarization could yield very human readable messages. However, it only picked portions of the original articles and failed to effectively convey the richness of the original notions; and missing some in the making. The study turned to abstractive summarization to take full advantage of deep learning language models. Due to hardware and time limitations, only 72,000 article-summery pairs were used. The resulting model with GPT, although not perfect, was the best and satisfactory measured with multiple ROUGE scores.
E-commerce live streaming was a new key battleground at the time of their writing. The logistics behind have a lot of complications; Key Opinion Leaders (KOLs) were only the front-end. Despite failures in the West, Multi-Channel Networks (MCNs) became important strategic marketing assets. MCNs managed KOLs and supported related dealings, from styling, sales platform management to product promotion contracts. For sales optimization, a strategic decision was on the influencer-platform-product mix. Tracking over 4000 KOLs, attributes on 4.3 million streaming sessions selling 4.6 million products were obtained. The students investigated the problem with thorough analyses, from basic descriptive statistics to hierarchical clustering in Python and various classification models with SAS Enterprise Miner. They found effective promotion strategies, and elaborated results with visualization of maps and word clouds.
Social media popularity skyrocketed during and after Covid-19 lockdowns in China. The cultural preferences also changed. This study investigated the changes of user sentiments based on regular comments and danmu on video sharing. The comment-rich media helped business understand their customers. The students collected over 2 billion live comments for over 86 thousand of videos before and during the pandemic (two full years of 2019-2020). While naive word counting could be misleading, as some words are naturally more frequently used, they deployed Jieba and Term Frequency-Inverse Document Frequency (TF-IDF) vectorization in Python as the feature extraction mean. On top of basic descriptive statistics presented with word clouds and usual charts, they used various classification models with SAS Enterprise Miner to uncover the patterns of changes. They helped promotion tactics by identifying effective and trendy words for titling, tags and keywords.
Cryptocurrencies have become one of the mainstream investment vehicles. However, notorious is their volatility. Traditional daily data unnecessarily distorted the actual trading information as a result of aggregation by time or other bins. Tick-to-tick data give the most truthful view of the market's trading information (e.g., the timestamp, the transaction type, the trading volume etc.). However, early econometric models were not designed for tick-to-tick volatility analysis upon irregular time intervals. Thus, this study aims to adopt a deep learning approach to assess the volatility measures for price prediction. New manual features were derived from the principles of close-to-close, Garman-Klass and Parkinson volatility. The cryptocurrencies under study were Cardano (ADA), Bitcoin (BTC), Dogecoin (DOGE), Polkadot (DOT), EOS, Ethereum (ETH), Litecoin (LTC), Ripple (XRP). Typical models of LSTM and GRUs were used. Results show that the manual features to deal with irregular time intervals and volatility can effectively help prediction performance. Ironically, the simple close-to-close volatility generally performs the best, based on the measure of mean absolute errors.
Value at Risk (VaR) has been widely used as a measure of market risk. It is considered a prudent statistical tool to quantify possible losses of an investment asset and has become a means for risk management in the mainstream. It is particularly important for the volatile market. However, the conventional methods were unable to handle the fluctuations in the market, especially in high-frequency trading. With the development of deep learning technology and the popularity of quantitative investment strategies, this study aims to validate the effectiveness of VaR with a backtesting strategy, and hence develop an effective automatic stock trading system. The key component is built upon deep reinforcement learning. With high-frequency data on 30 blue chips traded in the Hong Kong Stock Exchange, the students tested the 4 major deep reinforcement learning algorithms (SAC, PPO, A2C, and DDPG). Their model validated the effectiveness of VaR and found that SAC performed the best. Backtesting results showed that their developed automatic trading system could beat the bear market by almost 3% in 5 days on February 2022.
Cryptocurrencies have been accepted as digital assets in property law. However, notorious is its use as money-laundry vehicle. Anomaly detection methods have been around to fulfill anti-money laundry regulations. However, their trading volumes and frequencies invalidate the effectiveness. Daily or any aggregated data with traditional statistical econometric models were not able to identify anomalies under such conditions. Tick-to-tick data reflects the most truthful view of the market's trading information (e.g., actual transaction price, trading volume etc.). This study aims to adopt a deep learning approach to detect trading anomalies. Models were built and tested with basic principles in anomaly detection. Under study were five top cryptocurrencies in Binance including Bitcoin (BTC), Ethereum (ETH), BNB, Ripple (XRP), and Cardano (ADA). Typical models of LSTM and GRUs were used to analyze over 120 million transactions from 2020/01/01 to 2020/07/01. Results show that these strategies could supersede calendar anomalies and LSTM effectively narrowed down the detected anomalies in most cases.
Non-Fungible Tokens (NFTs) revolutionize the art industry as shareable collectibles with definite ownerships. They have become a major investment asset in the digital world; but like all other fine-art artifacts, valuation is difficult. To explore and facilitate the valuation process, this study attempted to estimate based on the contemporary market values of over 15,000 NFT collections. Collection characteristics like popularity, rarity and trading frequency were considered. With repeated use of decision trees on some meaningful variable transformations, it reveals that the floor prices of NFT collections have strange and non-linear complex relations with multiple characteristics. It provides valuable insights in estimating fair market prices of art pieces to researchers, investors, and NFT professionals.
The introduction of derivatives was supposed to stabilize the market, but instead, they have only added complexity. High-frequency crypto trading is all the more unpredictability. It is unsure if those principles that govern traditional financial markets still work in the crypto market. By leveraging derivatives metrics and tick-to-tick level transaction data, this study investigated market conditions like volatility, liquidity, market depth and etc. Typical deep learning models were used to analyze aggregated data of over 167 million data points with carefully engineered features. Results reconfirm some of the old principles and provides actionable insights to market regulators and market makers about pathways to identify key factors that affect market depth in the crypto derivative market.
The cryptocurrency market is notorious for its volatility, dramatic even in a short term. It challenges investors and traders to make informed decisions. Nevertheless, traditional wisdom in high finance says liquidity plays a crucial role in predicting volatility in the present of futures. This study aims to analyze some unique phenomena in the cryptocurrency market. It used tick-to-tick transaction data from Binance exchange and engineered 19 trading features for some deep learning analyses. The phenomena were reconfirmed. Results help smart money to enter or exit the market at a reasonable price; and also provide insights and references for regulators to stabilize the market and help confidence in the cryptocurrency market.
This study investigates how Natural Language Processing (NLP) can be harnessed to enhance educational outcomes by guiding large language models (LLMs) through pedagogically-informed prompt engineering. Focusing on the state-of-the-art LLM model, the project aims to teach LLM to answer examination True/False questions. Without the students teaching, the LLM can only achieve about 72.62 % accuracy. However, after their education-theory backed teaching, an LLM can achieve a marked performance lift to over 80.12 % accuracy using few-shot learning. They illustrate the power of instructional design in shaping LLM reasoning. These findings suggest that LLMs can emulate human-like learning when exposed to well-structured, theory-driven prompts. Their work demonstrates that prompt engineering can serve as a bridge between AI technology and instructional design: to offer educators and developers a pathway to create more accurate, explainable, and learner-centered AI tutors.
Cryptocurrency markets have evolved into pillars of digital finance, yet their extreme volatility remains difficult to model with traditional econometrics. This study investigates non-linear volatility transmission between BTCUSDT and ETHUSDT perpetual futures using high-frequency Binance tick data. We deploy a multivariate masked Long Short-Term Memory (LSTM) network alongside a Causal LSTM framework. Its inclusions are rich with order book microstructure features such as bid-ask spread, order imbalance, and market depth. By intelligently masking missing observations rather than applying naive imputation, the model captures true market dynamics, achieving an R² of 0.9874 for ETHUSDT on unseen data. Our results confirm that liquidity metrics, particularly the bid-ask spread, are powerful volatility predictors. Contrary to conventional assumptions of direct Bitcoin dominance, we find that ETH's short-term volatility is primarily driven by its own liquidity dynamics rather than cross-market contagion. This demonstrates that microstructure-aware deep learning significantly outperforms traditional approaches in capturing real-time crypto volatility.
Stablecoins are presumably stable, but investors may be overconfident in that stability. Previous studies relied on linear models and low-frequency data, yielding mixed results on USDT-BTC causality. This project re-examined the relationship using high-frequency data and a non-linear deep learning approach, asking whether stablecoin activity predicts Bitcoin's trading environment, whether a direct short-term causal link exists. The analysis integrates price dynamics like volatility, liquidity, and market microstructure signals like order flow imbalance to capture the granular mechanics of crypto markets. By using advanced Deep Learning model, the study provides a richer, more realistic picture of how stablecoins and volatile cryptocurrencies co-evolve-and whether that stability investors count on is as reliable as it means to be.
Financial fraud not only erodes investor trust but also distorts capital markets by mispricing assets and misallocating resources. Detecting it is notoriously difficult: fraud is rare, rapidly evolving, and costly to misclassify, meaning algorithm choice drastically shapes outcomes. This research followed the shift from traditional statistical tests to machine learning, where accuracy, stability, and interpretability now drive real-world deployment. Based on principles of the Fraud Triangle, it specified the variables to include. It confirmed that some pre-processing requirement in the literature is still valid in even the most power machine learning predictive models. Crucially, the study demonstrates that financial theory and model selection are both important in evaluating high-risk sample and in auditing. This project shows how data-driven methods can transform fraud detection from a theoretical exercise into a practical, business-ready tool.
Convertible bonds blend debt protection with equity upside, creating complex risk-return dynamics driven by underlying stock prices, conversion premiums, maturity, and embedded options. From an investment perspective, an issue is not the price level, but the drawdown risk and return opportunity over a future holding period. This project develops a two-stage machine learning framework for the Chinese convertible bond market to systematically navigate this complexity. Stage one screens the universe for a lower-risk safety pool based on drawdown exposure, while stage two ranks assets within that pool using excess-return predictions. Comparing multiple algorithms from machine learning to deep learning models like LSTM and Transformer, the best one generates a robust portfolio delivering a 19.77% annualized return, 1.23 Sharpe ratio, and just -9.20% maximum drawdown over 69 months. In the same period, the HS300 benchmark returned of 1.84% with a -35.95% drawdown. These results prove such integrated system works satisfactory.
In the world of data analytics, there is a professional examination recognized worldwide. Some of our students headed for that and acquired that professional qualitification. They all got very good starts at the early career.
Take minutes to guide progress (your supervisor may have other template, follow theirs if so)
On research
On project
M534, Faculty of Business, Meng Tak Building, Macao Polytechnic University, Rua de Luís Gonzaga Gomes, Macau
Tel: (853) 8599 3312
Fax: (853) 2872 7653
Email: billyyu@mpu.edu.mo
(check out the Google Scholar list, that is more updated)
Note:
The intended purpose of my Facebook Messenger was for leisure chatting and non-academic development. Nevertheless, not much actual demand for what I intended. Instead, forgive me, I am too slow to recognize that I might have invited possible unfairness and misunderstandings that students may skip classes, not take notes on classes, be forgetful, not read course materials, not read books and not discuss with their peers, etc. Some young students use instant messaging to ask for personal privileges, like examination tips and extra private tutoring. It must be too tempting.
However, I should be fair to all and give no favor to individual requests. Such questions should be raised in class while all can benefit without partiality. Thus, to minimize conflict of interests, I would delink students in my current teaching courses and refuse friend invitations for them during that semester. We can meet face-to-face regularly in class anyway. The official channels are good enough. Of course, after the course, you may add me again as your friend if you choose to.
Enjoy studying, and be a real university student! Have fun!