A Multiphase Deep Learning Framework for Stock Fundamental Analysis

다국어 초록 (Multilingual Abstract)

This dissertation proposes a novel, multiphase deep learning framework for fundamentals-based stock selection, designed to estimate fundamentals-implied stock values and identify undervalued firms grounded in value investing principles. The model integrates both temporal and cross-sectional financial perspectives to generate scalable, interpretable signals using firm-level data alone. The framework consists of three phases. Phase 1 extracts latent features through two representation learning models: a Temporal Fusion Transformer (TFT) processes time-series financial statement data, while an attention-based autoencoder encodes peer-relative structures from financial ratios. Phase 2 combines these embeddings to predict next-period stock values, which serve as fundamentals-implied values. In Phase 3, undervalued stocks are identified by comparing predicted and actual prices, and a composite score incorporating value premium, the difference between predicted and actual prices of undervalued stocks, and financial health is used to construct 20-stock portfolios. Empirical validation across four economic regimes (crisis, bull market, COVID-19, and recent period) and market-cap universes (small-, mid-, large-, and combined-cap) shows the model consistently outperforms benchmarks such as the S&P 500 and value ETFs in Sharpe ratio, Sortino ratio, and downside risk metrics. Feature representation analysis confirms the complementarity of temporal and cross-sectional inputs, while ablation tests, placebo experiments, and rolling window analyses affirm the model’s robustness and financial interpretability. This study demonstrates that deep learning can operationalise valuation logic, estimate intrinsic value proxies, and construct fundamentals-aligned portfolios in a scalable manner. It offers both methodological and practical contributions to the intersection of machine learning and fundamental investing.

목차 (Table of Contents)

• Chapter 1. Introduction 1
• Chapter 2. Data Characteristics and Analytical Rationale 10
• 2.1 Financial data description 10
• 2.2 Analytical and theoretical rationale for time-series 12
• 2.3 Analytical and theoretical rationale for cross-section 17
• 2.4 Combination of time-series and cross-sectional information 22
• Chapter 3. Proposed Multiphase Deep Learning Model 27
• 3.1 Model overview 28
• 3.2 Phase 1: Feature extraction 32
• 3.3 Phase 2: Estimation (prediction) 46
• 3.4 Phase 3: Stock selection based on value investing 51
• Chapter 4. Data preparation and simulation setup 57
• 4.1 Backtesting design 57
• 4.2 Training strategy 64
• 4.3 Portfolio simulation setup 69
• 4.4 Portfolio metrics 73
• 4.5 Placebo test setup 77
• Chapter 5. Simulation Results and Analysis 79
• 5.1 Phase 1 latent representative features insights 79
• 5.2 Model design validation 88
• 5.3 Stocks selected by the multiphase model 94
• 5.4 Portfolio metric comparison 101
• 5.5 Additional portfolio metrics 104
• 5.6 Placebo test results 107
• 5.7 Alternative stock selection strategy comparisons 109
• Chapter 6. Conclusion and future studies 114
• References 119