Analysis of the Korean stock market and portfolio optimization model using Autoencoder

다국어 초록 (Multilingual Abstract)

We analyze the Korean market，as it has the unique feature of having a large proportion of companies with the largest market capitalization relative to the total market. Because this has a huge impact on the market，we effectively analyze the market through major factors and better portfolio construction to explore the differences between the Korean and US markets．The market analysis is performed using factors，an important concept in financial markets，and factors extracted by Autoencoder，which is effective in reducing dimensionality．The correlation analysis between the factors and the market data suggests that the Korean market is more difficult to explain, both linearly and nonlinearly, than the US market. We then compare the performance of a model that uses the return data predicted by the Autoencoder，which is effective at denoising，as input to a portfolio optimization problem，to a traditional model. We further analyze focus on largest market capitalization stock to find the specificity of the Korean market．While the research confirms the performance of Stacked Autoencoder in terms of market analysis，denoising, and investment performance of the portfolio using Stacked Autoencoder，it does not find any significant differences between the Korean and US markets for largest market capitalization stock in the composition of the portfolio．Therefore，it suggests that further research is needed to consider additional factors beyond a simple approach such as including or excluding largest market capitalization companies in the portfolio composition，and to consider data other than returns．

목차 (Table of Contents)

• 1. Introduction 1
• 1-1. Background and Purpose 1
• 1-2. Literature review 3
• 1-3. Framework 4
• 2. Analysis methodology 5
• 2-1. Data 5
• 2-2. Principal components analysis 6
• 2-3. Autoencoder 6
• 2-4. Performance metrics 6
• 2-5. Permutation importance 7
• 2-6. Portfolio Optimization 8
• 3. Analysis results 9
• 3-1. Results from principal components analysis 9
• 3-2. Results from Stacked Autoencoder 10
• 3-3. Correlation analysis 14
• 3-4. Results and Implications 16
• 4. Empirical analysis of portfolio optimization 18
• 4-1. Stacked Autoencoder GMV 18
• 4-1-1. Backtest Structure 18
• 4-1-2. Compare Variances 18
• 4-1-3. Investment Performance 19
• 4-1-4. Results and Implications 24
• 4-2. MC1 Analysis 25
• 4-2-1. Investment weight 25
• 4-2-2. SAE Error 28
• 4-2-3. Correlating MC1 error rank and cumulative return 31
• 4-2-4. Results and Implications 31
• 4-3. Rank Portfolio Optimization 33
• 4-3-1. Investment Performance 33
• 4-3-2. MC1 Correlation Analysis 37
• 4-3-3. Tracking error 38
• 4-3-4. Results and Implications 39
• 5. Conclusion 41
• 6. References 42