A Conceptual Facet-based Framework Facilitating Transparent Design of Deep Learning Models in Financial Forecasting

Financial time series forecasting remains a challenging task due to the influence of numerous complex and interdependent internal and external factors. While deep learning models have shown promising performance in this domain, their design processes often lack transparency, with critical decisions, such as market selection, company scope, data coverage, model complexity, and use of technical indicators, being made without clear justification. This paper proposes a conceptual facet-based framework to guide the transparent and reproducible design of deep learning models for stock price prediction. By systematically defining five key design facets - market choice, company size, data time span, model complexity, and technical indicator usage - this framework enhances transparency and supports more rigorous experimentation. Extensive empirical studies across companies in the Mainland China and Hong Kong markets demonstrate the significant impact of each facet on model performance. The findings provide actionable insights for improving predictive accuracy and reproducibility, while laying a foundation for more trustworthy financial AI systems.