# ETF Portfolio Construction via Neural Network trained on Financial Statement Data

**Authors:** Jinho Lee, Sungwoo Park, Jungyu Ahn, Jonghun Kwak  
**arXiv:** [2207.01187](https://arxiv.org/abs/2207.01187)  
**Date:** 2022-07-04  
**Summary page:** https://www.summarizepaper.com/en/paper/2207.01187v1/

## Key points

- Application of advanced machine learning methods for asset management has gained attention
- Data shortage problem hinders utilization of deep neural networks
- Novel approach using neural networks to construct ETF portfolios based on financial statement data proposed
- Leveraging individual stock data to train neural networks for predicting future performance
- Proposed method outperforms existing baselines in managing ETF portfolios
- Particularly beneficial for managing recently listed ETFs with limited historical data
- Previous research focused on technical and fundamental analysis, as well as alternative data sources like investor sentiment
- Combining neural networks with financial statement data offers a promising solution to address data scarcity in asset management

## Summary

In recent years, the application of advanced machine learning methods for asset management has garnered significant attention. However, the utilization of these methods, particularly deep neural networks, is hindered by the data shortage problem. To tackle this challenge, a novel approach using neural networks to construct a portfolio of exchange traded funds (ETFs) based on the financial statement data of their components has been proposed. While ETFs and ETF-managed portfolios have gained prominence in recent decades, applying neural networks to manage ETF portfolios is constrained by the relatively smaller number and shorter historical existence of ETFs compared to individual stocks. To overcome this limitation, the data of individual stocks is leveraged to train neural networks for predicting the future performance of individual stocks. These predictions, along with the portfolio deposit file (PDF), are then used to construct a portfolio of ETFs. Through multiple experiments, it has been demonstrated that this proposed method outperforms existing baselines. This approach is particularly beneficial when managing recently listed ETFs like thematic ETFs, which have limited historical data available for training advanced machine learning models. Furthermore, previous research efforts in stock market predictability have primarily focused on technical analysis-based approaches that rely on historical price and volume data or fundamental analysis-based approaches utilizing financial statement data from individual stocks. Alternative data sources such as investor sentiment in social media messages or news articles have also been explored to identify profitable signals for trading. Overall, this innovative approach combining neural networks with financial statement data for constructing ETF portfolios presents a promising solution to address the challenges posed by data scarcity in asset management. By leveraging insights from individual stock data and predictive modeling techniques, this methodology offers a robust framework for managing ETF portfolios effectively in dynamic market conditions.