The paper 'Exploring Candlesticks and Multi-Time Windows for Forecasting Stock-Index Movements', written by Seo Kang-hyun (combined master's–doctoral student and first author) of the Machine Learning Laboratory in the Department of Computer Science & Engineering and corresponding author Professor Yang Ji-hoon (supervisor), has been accepted for publication at the ACM/SIGAPP Symposium on Applied Computing (SAC) — Machine Learning and Its Applications Track, listed at an adjusted IF of 1 among the outstanding international conferences in computer science under the BK21 Plus programme.

▲ (From left) Seo Kang-hyun (combined master's–doctoral student) and Professor Yang Ji-hoon (supervisor)

The aim of the research is to predict through binary classification whether the closing value of a composite stock index has risen or fallen relative to the previous day; the proposed model is named Combined Time View TabNet (CTV-TabNet).

CTV-TabNet applies 'TabNet: Attentive Interpretable Tabular Learning', published by Google in 2019, to learn the features of each candlestick chart shape, and then uses GRUs, which perform strongly on time-series data, to learn from input data of various lengths. The input data is then divided by length into microscopic and macroscopic views, and a neural network combines the two views into a single final prediction.

The reason for dividing the input data in various ways in the GRU is that stock price movements are continually influenced by multiple momenta, and the aim is to reflect these momenta in various ways.

CTV-TabNet not only showed higher prediction accuracy than the comparison models across 20 composite stock index datasets but also achieved higher returns than the comparison models in a futures trading simulation.

The ACM/SIGAPP SAC conference is a distinguished international conference covering a wide range of computing fields. It will be held in Tallinn, Estonia, from 27 to 31 March next year.

SAC 2023 link: https://www.sigapp.org/sac/sac2023/index.html