Core Insights - The report studies a paper titled "(Re-)Imag(in)ing Price Trends," which presents a method for K-line image recognition and trend prediction based on convolutional neural networks (CNN), aiming to localize this approach for the domestic market [1] Group 1: Research Background - The paper automates the visual analysis process of K-line charts, addressing limitations in traditional financial models that rely on subjective human experience [11][14] - The innovative approach utilizes machine learning to discover predictive patterns from data without pre-setting specific models, aligning more closely with how traders analyze charts [11][14] Group 2: Model Essence - The first step involves generating standardized K-line technical charts from historical market data, utilizing daily frequency data from the CRSP database covering 1993-2019 [11][12] - The CNN model is designed to automatically extract local features through convolution and pooling operations, with a focus on predicting future return directions rather than precise values [14][18] Group 3: Empirical Results - The model demonstrates strong predictive accuracy, achieving a 53.3% accuracy rate for predicting 20-day returns, significantly outperforming random guessing [19][20] - In portfolio construction, a long-short strategy based on 20-day images yields an annualized Sharpe ratio of 2.2, far exceeding traditional momentum strategies [22][24] Group 4: Practical Application - The model's transferability is validated, showing that a model trained on U.S. stocks can be applied to 26 other countries, often outperforming locally trained models [25][28] - Initial applications in the domestic market using data from 20 major ETFs since 2020 achieved a classification accuracy of 55.3%, indicating the model's ability to extract valuable information from K-line images [37][39] Group 5: Investment Practice - The report proposes a localized model construction process, emphasizing the importance of data diversity to avoid overfitting and enhance the model's learning capabilities [35][36] - The model's design includes data cleaning, standardization, and the generation of 2D images from raw price-volume data, followed by training using a deep learning framework [36][37]