This paper presents a deep learning-based investment decision strategy that transforms raw stock price time series into candlestick chart images, which are then processed by a Convolutional AutoEncoder (CAE) to learn compact, nonlinear feature representations. The learned features capture semantic information such as industrial sector and address limitations of traditional similarity measures like linearity, long data requirements, and lack of translation invariance. Using these features, the method applies modularity optimization to cluster stocks into diverse groups without predefining the number of clusters, ensuring deterministic and data-driven market segmentation. Portfolios are constructed by selecting the stock with the highest Sharpe ratio from each cluster and allocating equal weight to each selected stock.

Experimental evaluation on FTSE 100 data over 4000 trading days demonstrates that the deep feature-based approach significantly outperforms raw time series features and alternative clustering methods like k-means. The proposed strategy achieves a total return of 283.5% over the long term, surpassing the FTSE 100 index and several well-known funds (e.g., CCA, VXX, IEO, PXE, PXI) in total return, daily Sharpe ratio, and other risk-adjusted metrics. The method also shows robustness across different market conditions, including bearish, flat, and bullish periods. The work contributes a novel deep learning signal for portfolio optimization and highlights the effectiveness of combining visual representation learning with modularity-based clustering for investment decisions.