Quant Crunch

Team members

Lucas Ng Yi Leang (ESD), Ng Jing Da (ESD), Evan Sidhi Peradijaya (ESD), Tan Yin Ling (ESD), Chew Wen Jie Jeremy (ISTD), Koh Ting Yew (ISTD), Leo Ding Hao (ISTD)

Instructors:

Kenny Choo, Wang Xingyin, Lynette Cheah

Writing Instructors:

Grace Kong

Teaching Assistant:

Cheong Rui Zhi Jeremy, Anirudh Gajendra Rathi

Project Motivation

Part of the energy that drives the project lies in the inherent passion the team has as retail traders. From the get go, the team was invested into making an algorithmic platform that seeks to simplify the onboarding process as well as deliver the assurance that users need when adopting a new technology.

Through a series of surveys and user testing evaluations, we found out that newbie traders were often bombarded with a huge amount of information. This led to Quant Crunch which is an algorithmic tool leverages upon the use of Deep Learning to provide trade recommendations that enables users to acquire an edge in their existing trading strategies.

User Requirements

User testing analysis was conducted by performing user observations studies to understand the general trading workflow, as well as to gather user requirements for the system. The process was conducted through Zoom where participants would share their webcam and screen with the interviewer so that their activities could be recorded. Any additional tools or systems that the participants used and their specific functionalities and purposes were also recorded and noted down.

Upon completion of the study, the notes were consolidated and a simple affinity mapping was made to identify recurring themes and patterns. The plan was to use the qualitative data that the team obtained to generate insights which could be used to guide the design process for the user interface.

Needs of our Target Audience:

1) Ease of Use

2) Appropriate Risk-to-Return Performance

Model Considerations

The team tested various approaches and algorithms, and this is a top level overview to how we designed and evaluated our models.

The model can either forecast how the price might change or decide what the best action would be at a given time.

In terms of price prediction, the team tried out a neural network based model, which used a recurrent network architecture, as those have been shown to be useful when dealing with sequential data. We also looked into technical analysis indicators, which are conventional trading disciplines to help identify trends and patterns seen within price data. The use of TA helped to address one of the issues found when dealing with price forecasting models, which is the penchant to produce shifted outputs.

The team also looked at action-predictive models, including a simpler version of the price-predictive model that uses a binary classifier to decide which of buy or sell would be most appropriate at a given time, as well as a reinforcement learning model which uses trial and error to explore the trading environment and decide which action maximises its cumulative reward.