# Financial Time Series Forecasting using CNN and Transformer

**Authors:** Zhen Zeng, Rachneet Kaur, Suchetha Siddagangappa, Saba Rahimi, Tucker Balch, Manuela Veloso  
**arXiv:** [2304.04912](https://arxiv.org/abs/2304.04912)  
**Date:** 2023-04-11  
**Summary page:** https://www.summarizepaper.com/en/paper/2304.04912v1/

## Key points

- Authors address the importance of time series forecasting in decision-making across various domains
- Challenges in predicting financial time series data due to modeling short-term and long-term temporal dependencies
- CNNs excel at capturing local patterns for short-term dependencies but struggle with long-term dependencies
- Transformers are effective at understanding global context and long-term dependencies
- Proposed novel approach combines CNNs and Transformers to model both short-term and long-term dependencies within a time series
- Goal is to forecast stock price movements (up, down, or flat) in the future
- Success demonstrated through experiments comparing their method with other approaches for forecasting intraday stock price changes for S&P 500 constituents
- Research contributes valuable insights into leveraging advanced neural network architectures for more accurate financial time series forecasting

## Summary

In their paper titled "Financial Time Series Forecasting using CNN and Transformer," authors Zhen Zeng, Rachneet Kaur, Suchetha Siddagangappa, Saba Rahimi, Tucker Balch, and Manuela Veloso address the importance of time series forecasting in various domains for decision-making. Specifically focusing on financial time series such as stock prices, the authors highlight the challenges associated with predicting these data points due to the complexity of modeling short-term and long-term temporal dependencies. While Convolutional Neural Networks (CNN) excel at capturing local patterns for modeling short-term dependencies, they fall short in learning long-term dependencies due to their limited receptive field. On the other hand, Transformers are adept at understanding global context and long-term dependencies. To overcome these limitations and harness the strengths of both CNNs and Transformers, the authors propose a novel approach that combines the two techniques to model both short-term and long-term dependencies within a time series. The goal is to forecast whether stock prices will go up, down or remain flat in the future. Through a series of experiments comparing their method with commonly adopted statistical and deep learning approaches, the authors demonstrate the success of their proposed model in forecasting intraday stock price changes for S&P 500 constituents. Published at AAAI 2023 - AI for Financial Services Bridge conference,this research contributes valuable insights into leveraging advanced neural network architectures for more accurate and reliable financial time series forecasting. By integrating CNNs' ability to capture local patterns with Transformers' capacity to understand global context,this hybrid approach offers a promising solution for addressing the challenges associated with predicting stock prices accurately over different time horizons.