# Predicting Stock Price Movement as an Image Classification Problem

**Authors:** Matej Steinbacher  
**arXiv:** [2303.01111](https://arxiv.org/abs/2303.01111)  
**Date:** 2023-03-02  
**Summary page:** https://www.summarizepaper.com/en/paper/2303.01111v1/

## Key points

- The paper uses a CNN-based model to explore intraday price movement of stocks as an image classification problem.
- The study shows a high-level relationship between the first hour of trading and the close, with the algorithm separating opposing classes effectively.
- Investing according to the algorithm's predictions outperformed all alternative constructs except for the theoretical maximum.
- Several additional tests were conducted to support their thesis, highlighting the suitability of computer vision techniques for studying financial markets and predicting stock price movements.
- The study addresses overfitting by adding a dropout layer that discards 20% of random units prior to classification.
- Various financial and economic variables proposed as potential predictors are discussed, including valuation ratios, interest rates and spreads, macroeconomic variables such as inflation and industrial production, and psychology in asset pricing.
- Recent advancements in artificial intelligence have entered the field of financial markets, with some examples combining pipelines of 1-dimensional CNN models with bi-directional LSTM units or using CNN models on limit order books.
- Asset-trading can be considered an image classification problem done on an in-sample dataset.
- Technological innovation may diminish the size of investment effects while emphasizing that a smaller portfolio is more manageable.

## Summary

The paper explores the intraday price movement of stocks as an image classification problem using a CNN-based model. The study demonstrates a high-level relationship between the first hour of trading and the close, with the algorithm adequately separating between opposing classes. Investing according to the algorithm's predictions outperformed all alternative constructs except for the theoretical maximum. To support their thesis, several additional tests were conducted, highlighting the suitability of computer vision techniques for studying financial markets and predicting stock price movements. The study also addresses overfitting by adding a dropout layer that discards 20% of random units prior to classification. The paper discusses various financial and economic variables proposed as potential predictors, including valuation ratios, interest rates and spreads, macroeconomic variables such as inflation and industrial production, and psychology in asset pricing. While these papers contributed to fundamental understanding, they had limited use for predicting individual stock price movements. Recent advancements in artificial intelligence have entered the field of financial markets, with some examples combining pipelines of 1-dimensional CNN models with bi-directional LSTM units or using CNN models on limit order books. The paper concludes that asset-trading can be considered an image classification problem done on an in-sample dataset. The author suggests that technological innovation may diminish the size of investment effects while emphasizing that a smaller portfolio is more manageable. Matej Steinbacher graduated in economics from the University of Maribor and continued post-graduate studies at the Faculty of Economics at the University of Ljubljana focused on network models in economics and finance.