Data Science Portfolio

Education

Proposed a novel model integrating graph discovery data mining techniques and Graph Neural Networks(GNNs) to overcome limitations in graph structures for specific modeling tasks. By incorporating graph discovery models, the need for domain expertise in graph generation was also eliminated, resulting in significant time and effort savings compared to manual graph creation.
It demonstrated a minimum of 15% improvement of the baseline Light Gradient Boosting Machine (LightGBM) and Neural Net models in a short-term intraday Realized Volatility (RV) prediction task by using 150 million records Limit Order Book (LOB) and trade data.

Created SmoothLaplacian Python network analysis library, facilitating efficient underlying network discovery of diverse data. This tool streamlined the process of extracting network features from complex datasets.
It is a Python implementation of Learning Laplacian Matrix in Smooth Graph Signal Representations

Developed a Graph Neural Networks (GNNs) based recommender system connecting businesses with social media influencers by learning various aspects of their social media behaviors such as text caption, image, and the social relationship among influencers and brands.
Specifically, the model obtained the representations of businesses and social media influencers by Light Graph Convolutional Networks (LightGCNs).
Collected and cleaned a dataset comprising 50GB pool of raw user profiles and posts collected from over 1427 businesses, and 16774 influencers.
Extracted features from social media user behaviors by employing pretrained Natural Language Processing (NLP) and computer vision (CV) models.

This is a network consisting of selected brands and influencers from our dataset. This is an interactive version of this network.

Produced the stock network discovery dashboard that allows one to discover the latent network within any selection of stocks during any period of time.

Used a Simulated Minimum Distance (SMD) estimator based on the Wasserstein distance, Wasserstein Distance Estimator (WDE), between the model-simulated distribution and the empirical distribution.
The method resolved common challenges posed by intractable analytical likelihood functions by implementing the Wasserstein distance estimator for complicated time series models, ensuring accurate estimations where Maximum Likelihood Estimator (MLE) fell short.
The method outperformed existing estimation methods, both in terms of accuracy and computational time (2 - 100x faster). Models estimated include AR(1) model, ARMA(2, 2)-ARCH(2), random walks with a structural break and 10 years S&P 500 data.

Instructor for courses including Statistics and Linear Algebra. Materials include experimental design, hypothesis test, dimension reduction, data science application, and Python demonstrations. An example.

Provided support to internal cross-functional teams and external ETF fund managers through data collection, statistical analysis, and modeling, with a primary focus on summarizing index behaviors, including summary statistics and linear regression.
Researched, developed, and presented the Canada Dividend Achiever Index report including index description, calculation, evaluation, maintenance, and rebalancing, tailored for sales initiatives and investors’ comprehensive analysis.

Conducted research reports on investment opportunities in Fund of Funds (FoFs) for investment of funds.
Contributed to the development and deployment of high-quality accounting automation to expedite delivery and improve efficiency.