Multi-Timeframe Transformer Architecture for Bitcoin Price Prediction: Deep Dive into the Research Project

Gandrathi, Sanjana (2025) Multi-Timeframe Transformer Architecture for Bitcoin Price Prediction: Deep Dive into the Research Project. Masters thesis, Dublin, National College of Ireland.

Preview	PDF (Master of Science) Download (873kB) | Preview
Preview	PDF (Configuration Manual) Download (4MB) | Preview

Abstract

The cryptocurrency markets are highly volatile and have complicated temporal dynamics that make conventional forecasting techniques less effective. Although deep learning methods have been promising, they generally study individual timeframes and do not adequately combine sentiment signals. This research investigates whether a multi-time-frame Transformer architecture can improve Bitcoin price prediction accuracy compared to traditional LSTM models, while examining how social sentiment importance varies across temporal resolutions.

The study develops a novel Transformer architecture that concurrently processes hourly, daily, and weekly temporal patterns through specialized cross-timeframe attention mechanisms. The model integrates 97,000 sentiment-labeled social media messages with historical price data spanning 2014-2019, creating a comprehensive dataset of 41,120 samples with 104 engineered features. Comparative experiments evaluate five model configurations including single and multi-timeframe variants with sentiment ablation studies.

With multi-timeframe LSTM attaining a 34.8% RMSE reduction over single-time-frame variants, the results show that the multi-timeframe approach offers significant improvements. But surprisingly, the multi-timeframe LSTM performs better than the Transformer architecture, attaining 57.8% directional accuracy as opposed to 56.6% for the Transformer. Interestingly, sentiment integration shows model-dependent effects, reducing LSTM performance by 7.7% while increasing Transformer accuracy by 20.7% RMSE reduction. With a directional accuracy of 62.1%, the LSTM architecture demonstrates exceptional strength in 24-hour predictions. These results imply that while architectural advancements in temporal processing for cryptocurrency forecasting hold promise, they primarily depend on matching architecture to data attributes. With implications for trading systems and risk management protocols, the paper presents a well-established paradigm for analysis across multiple timescales and reveals intricate relationships between model architecture and sentiment feature success.

Item Type:	Thesis (Masters)
Supervisors:	Name Email Del Rosal, Victor UNSPECIFIED
Subjects:	H Social Sciences > HD Industries. Land use. Labor > HD61 Risk Management H Social Sciences > HG Finance > Money > Digital currency > Cryptocurrencies H Social Sciences > HG Finance > Fintech T Technology > T Technology (General) > Information Technology > Fintech
Divisions:	School of Computing > Master of Science in FinTech
Depositing User:	Ciara O'Brien
Date Deposited:	20 Aug 2026 10:45
Last Modified:	20 Aug 2026 11:36
URI:	https://norma.ncirl.ie/id/eprint/9569

Actions (login required)

View Item