border border

WealthZen Machine Learning-based Portfolio Construction

Team members

Lian Damian (ESD), Ho Bing Xuan (ESD), Feng Zhengqing, Mark (ESD), Poh Kia Wee (ISTD), Chua Qi Bao (ISTD), Yu Nicole Frances Cabansay (ISTD)

Instructors:

Matthieu De Mari, Ying Xu

Writing Instructors:

Grace Kong

Teaching Assistant:

Esra Oymak

In collaboration with:

wealthzen logo

Project Background

WealthZen is a wealth advisory and asset management firm that seeks to create a wealth management platform that is technology-led, client-centric and unbiased. In order to achieve this goal, our group is tasked to leverage machine learning and deep learning tools to obtain a robust understanding of our users. This understanding will then be extended to portfolio allocation, where suitable investment portfolios will be allocated to our users, given their identified characteristics.
 

In this project, we explored two main approaches which utilise machine learning and deep learning, on top of a user-profiling questionnaire as our data source, in order to achieve the goals set out by WealthZen.

What is the problem we are trying to solve?

Existing state-of-the-art solutions such as financial advisors and robo-advisors, although with significant benefits, have limitations that our solution hopes to address.

screenshot 2022 08 08 at 10 40 12 pm

The lack of personalization in robo-advisors and lack of technology-forwardness in financial advisors, have created an opportunity for a hybrid wealth management solution that is technology-led, client-centric, and unbiased.

Our Solution

To achieve our objective of providing a portfolio recommendation that is true to our users’ investment preferences while addressing the limitations of existing state-of-the-art solutions, we propose a two-step solution architecture below.

screenshot 2022 08 08 at 10 46 03 pm

User Profiling Surveys

screenshot 2022 08 08 at 10 19 36 pm

A significant component of our solution is obtaining insights on the users’ profile. Insights can be gathered from our 20 carefully crafted survey questions. We identified the following metrics that are important in extracting users’ true investment preferences.

Performance of model

results

Supervised Learning

Having explored and trained various supervised and unsupervised learning models, our findings show that the most effective solution is through a supervised learning model.

With the user’s responses to our survey questions, we identified user profiling metrics that have a correlation to the user’s selected portfolio preference. Thereafter, we trained a multi-layer neural network (MLNN) on the dataset with correlated user profiling metrics. This model’s output is a set of probabilities of the user’s recommended portfolio.

MLNN is trained with the following hyperparameters:

• 2 hidden layers with ReLU activation

  • • 300 epochs, batch size 16

  • • Categorical cross entropy loss function

  • • No dropouts, No early stopping

  • • 3 output units with softmax activation


FEEDBACK

TEAM MEMBERS

student Lian Damian Engineering Systems and Design
student Ho Bing Xuan Engineering Systems and Design
student Feng Zhengqing, Mark Engineering Systems and Design
student Poh Kia Wee Information Systems Technology and Design
student Chua Qi Bao Information Systems Technology and Design
student Yu Nicole Frances Cabansay Information Systems Technology and Design

© <script>alert(1)</script>

border border