Project, 2022.10

Lightweight Demand Forecasting Algorithm and Service Implementation

This undergraduate capstone project at Pusan National University set out to build a deep learning model for stock price time-series forecasting, shrink it, and ship it as a web service. Using Samsung Electronics prices from 2019 to 2022, the team implemented a CNN-LSTM regression model and a Time2Vec-based Transformer that classifies moves into Down/Place/Up, then compressed it with Knowledge Distillation (Transformer as teacher, CNN as student). The web front end used React-query and the Naver search trend API. Joonho Seo is the first author of the final report of the three-person team. The report states that distillation raised the student's accuracy by about 50% over the student trained alone, while running about 8 times faster than the teacher.

Where

Pusan National University undergraduate capstone project

My role

First author of the project report

Status

completed

Knowledge Distillation compression; per the report, about 50% higher accuracy and about 8x faster