知の物理学研究センター / Institute for Physics of Intelligence (iπ)

【日時/Date】
2022年7月11日（月）17時00分～18時30分

【講演者/Speaker】
南　賢太郎/MINAMI Kentaro（株式会社 Preferred Networks）

【講演タイトル/Title】
Deep Learning for Financial Applications

【概要/Abstract】
Recent developments in deep learning techniques have motivated intensive research in deep-learning-based financial applications. However, deep learning faces several fundamental challenges arising from financial markets, such as low signal-to-noise ratios, non-stationarity, and limited training data. It is common wisdom among deep-learning practitioners that imposing appropriate “inductive biases” often improves the data efficiency of deep learning. For example, convolutional neural networks (CNNs) are “biased” toward learning shift-invariant solutions, which are known to be effective in processing image data. In the context of Finance-Deep-Learning, it would be beneficial to utilize financial domain knowledge (such as finance and economics theory, heuristics, and stylized facts in financial markets) as inductive biases for deep learning models.In this talk, I introduce our two recent papers that demonstrate how we can improve the efficiency of deep learning by leveraging financial domain knowledge. In our first paper [1] on building profitable trading strategies, we propose several techniques to incorporate financial domain knowledge into modern deep learning systems. In particular, we utilize portfolio theory, factor investing, and some stylized facts in financial markets. In our second paper [2] on derivative hedging, we propose a new neural network architecture that reflects the structure of the theoretical optimal solution of a stochastic control problem. We show that our proposed architecture improves the speed of convergence and the utility of the trained model.
[1] Imajo, et al. (2021). Deep Portfolio Optimization via Distributional Prediction of Residual Factors. In AAAI2021. https://arxiv.org/abs/2012.07245
[2] Imaki, et al. (2021). No-Transaction Band Network: A Neural Network Architecture for Efficient Deep Hedging. https://arxiv.org/abs/2103.01775