RIC-NN: A Robust Transferable Deep Learning Framework for Cross-sectional Investment Strategy

Data science and AI in FinTech: an overview

“…These tasks apply deep learning models, including basic deep neural networks and their recent developments, such as diversified deep neural mechanisms, architectures and networks (e.g., recurrent neural networks, graph neu-ral networks, neural language models like Transformer and BERT variants, image nets, attention networks, memory networks, adversarial learning, and autoencoders, etc. [43,46,55]; -deep learning of illegal, noncompliant, risky and fraudulent behaviors such as insider trading and market manipulation in capital markets and financial accounting and reporting fraud in financial services and statements [12]; and -federated learning models for privacy-preserving opendomain and whole-of-business financial applications and services [19,57].…”

Transfer Ranking in Finance: Applications to Cross-Sectional Momentum with Data Scarcity

“…ability to extract universal features by showing that it is able to successfully transfer learn on a target set of stocks that are not part of the original source training set. The work of [45] develop the RIC-NN (Rank Information Coefficient Neural Net) and demonstrate that the lower layers of the optimised model's weights in one region (represented by MSCI North America) can be used to initialise the same model for prediction in a different region (MSCI Asia Pacific). In the same vein but with a broader coverage, [38] propose QuantNet which fuses information from different markets with a global bottleneck shared across multiple autoencoder-based architecture; The model then employs a separate decoder for each market to generate market-specific trading strategies.…”

“…Unlike earlier studies applying transfer learning to the problems in finance, ours differ in a few key aspects. Firstly, while works such as [38,45] primarily aims to transfer knowledge, ours (along with [30,34]) is explicitly directed at alleviating the issue of overfitted models that surfaces when calibrating on limited samples. Despite numerous studies focused on applying machine learning techniques to cryptocurrency trading, few, if any, explore the specific use of transfer learning and examine its application over lower frequency settings (such as over weeks) -where the problem of data scarcity is further exacerbated.…”

LSTM-DDPG for Trading with Variable Positions

“…Imajo et al [ 7 ] propose a system for constructing portfolios with a spectral decomposition-based method to hedge out common market factors and a distributional prediction method based on deep neural networks incorporating financial inductive biases. Nakagawa et al [ 8 ] propose a principled stock return prediction framework called Ranked Information Coefficient Neural Network (RIC-NN) to alleviate the overfitting. The learning difficulties of initialization, the stopping of training models and the transfer among different markets have been addressed.…”