Financial time series forecasting remains a fundamental challenge in data-driven investment, as transforming price dynamics into structured representations that can be effectively learned by deep neural networks is inherently difficult. Traditional sequence-based models often fail to fully capture the spatial patterns embedded in financial data. To address this limitation, this study proposes a unified image-based framework that integrates representation learning, predictive modeling, and portfolio optimization. Specifically, financial time series are transformed into grid-structured image representations by redesigning the Grid Shape Context Descriptor (GSCD) to encode both structural patterns and return magnitude within a unified form. These representations are used to train a hybrid convolutional neural network–long short-term memory (CNN-LSTM) model for return prediction, and the resulting forecasts are incorporated as investor views within the Black--Litterman framework for portfolio construction. Experiments on exchange-traded funds (ETFs) across developed and emerging markets show that the proposed framework consistently outperforms both traditional time-series models and existing image-based approaches in predictive accuracy, while also delivering superior risk-adjusted portfolio performance. The results further demonstrate robust performance across diverse market conditions, highlighting the effectiveness of GSCD-based representations in linking financial forecasting with practical investment decisions.