财政与货币金融教研室学术讲座
报告题目:How does Convolutional Neural Network predict stock returns?
主 讲 人:曾 祺(墨尔本大学金融学 副教授)
主 持 人:由林青 (中山大学岭南学院 助理教授)
时 间:2025年2月25日(周二) 下午15:00
地 点:岭南堂汪道涵会议室(101)
语 言:中 文
主要内容:
We interpret the image-based asset pricing factors produced with convolutional neural network (CNN) by proposing a methodology through which the aggregate pattern learned by vision models can be visualized in a heat map. We show and quantify the CNN captured price and volume paths, including the level, the slope, the curvature, and, most importantly, the last day change in the difference between close price and high-low average. Four simple geometric measures can explain 31% (24%) of variation in predicted probability of positive weekly (monthly) returns, and 73% (48%) of variation in weekly (monthly) CNN returns. We further hypothesize and confirm that the machine can identify time series persistence and reversal patterns. Separating them enables the formation of long-short portfolios with better performances than does the existing CNN factor.
主讲人简介:

曾祺墨尔本大学金融学副教授,2003年毕业于普林斯顿大学沃顿商学院,研究兴趣包括公司金融、资产定价和绿色金融。研究成果发表于Management Science 和Journal of Risk and Insurance等期刊上。
欢迎师生参加!