A Multi-modal Deep Learning Approach for Predicting Dhaka Stock Exchange

2023 IEEE 13th Annual Computing and Communication Workshop and Conference (CCWC)

Md. Nabil Rahman Khan, Omor Al Tanim, Most. Sadia Salsabil, S.M. Raiyan Reza, Khan Md Hasib, Mohammad Shafiul Alam

IEEE CCWC 20232023Published

Table of contents

Abstract

DeepDse forecasts closing prices on the Dhaka Stock Exchange by combining a Transformer encoder with LSTM and GRU layers. It is trained on almost 2 million rows of DSE trading data from 2007 to 2022, enriched with gold, silver, oil and US dollar prices and a set of technical indicators. Across 11 stocks from different sectors, the combined model had the lowest RMSE of every architecture we tested.

DeepDse architecture: a Transformer encoder feeding LSTM and GRU layers, pooling and dense output layers.

Key findings

Lowest RMSE on all 11 stocks tested, against a stacked Transformer, GRU and LSTM on their own
Mean RMSE of 0.020, compared with 0.029 for the Transformer, 0.046 for GRU and 0.066 for LSTM
Gold, silver, oil and USD prices plus RSI, MACD, VWAP, Supertrend and AROON as extra inputs

Results

AB Bank

AMCL (PRAN)

Apex Tannery

BD Finance

Grameenphone

Padma Oil

Popular Life

Samorita

RAK Ceramics

GBB Power

Beximco

RMSE on the test set for each stock. Lower (further left) is better. Tap a model to hide it; the axis rescales.

RMSE on the test set for each stock (lower is better).
Stock	Stacked Transformer	GRU	LSTM	DeepDse (hybrid)
AB Bank	0.0465	0.0577	0.0659	0.0458
AMCL (PRAN)	0.0152	0.0682	0.0771	0.0112
Apex Tannery	0.0631	0.0697	0.0889	0.0195
BD Finance	0.0262	0.0410	0.0463	0.0183
Grameenphone	0.0338	0.0426	0.0942	0.0233
Padma Oil	0.0339	0.0644	0.0729	0.0137
Popular Life	0.0243	0.0528	0.0847	0.0231
Samorita	0.0205	0.0245	0.0275	0.0140
RAK Ceramics	0.0185	0.0319	0.0662	0.0176
GBB Power	0.0134	0.0167	0.0621	0.0133
Beximco	0.0229	0.0336	0.0431	0.0228

Citation

Published in 2023 IEEE 13th Annual Computing and Communication Workshop and Conference (CCWC). DOI: 10.1109/CCWC57344.2023.10099255

View on IEEE Xplore Read the blog post

Deep Learning
Transformers
Time-Series Forecasting
Financial ML
Multi-modal Data

Co-authored by Omor Al Tanim

Software Engineer in Dhaka, Bangladesh. Skilled in AI agents, automation bots, and OCR solutions for intelligent process automation.

View profile

Comments (0)

Comments are checked by AI before they appear. Read the rules

Be respectful: no insults, harassment, hate or threats. Disagreeing is fine.
No spam: no ads, self-promotion, SEO or affiliate links, crypto or gambling.
Nothing sexual, violent or illegal, and no scams or malware.
Don't share anyone's personal details (phone numbers, addresses, private emails).
Keep it a real comment: no gibberish, test text or messages aimed at the AI.

Loading comments…