Loading...
|
Please use this identifier to cite or link to this item:
https://nccur.lib.nccu.edu.tw/handle/140.119/158594
|
| Title: | 多因子組合方法於台灣市場之實證研究：傳統動能、 降維與機器學習的綜合評估
An Empirical Study of Multi-Factor Combination Methods in the Taiwan Market: A Comprehensive Evaluation of Traditional Momentum, Dimensionality Reduction, and Machine Learning Approaches |
| Authors: | 陳昇華
Chen, Sheng-Hua |
| Contributors: | 林士貴
Lin, Shih-Kuei
陳昇華
Sheng-Hua Chen |
| Keywords: | 多因子
機器學習
動能
降維
交易策略
因子合成
Multi-Factor
Momentum
Dimensionality Reduction
Machine Learning
Factor Combination
Trading Strategy |
| Date: | 2025 |
| Issue Date: | 2025-08-04 14:33:19 (UTC+8) |
| Abstract: | 本研究檢驗 2010 至 2025 年間 52 項涵蓋估值、成長、獲利、品質、技術面與流動性維度的公司層級因子在臺灣股市的橫斷面報酬預測能力，資料取自 TEJ，並透過 MAD 截尾、規模與產業中性化及 Z 分數標準化三道程序處理，以確保訊號穩健且可比較。單因子分析顯示，價值與品質因子的資訊係數及資訊比率表現最佳，而動能、規模與風險導向因子波動較大。為整合多因子訊息，我們比較等權重組合、PCA、橫斷面與時間序列因子動能，以及 CatBoost、XGBoost 與 LightGBM 等梯度提升排序模型，結果以 LightGBM 最優，全市場樣本期間年化報酬率 15.90\%、夏普比率 2.50、最大回撤 $-5.90$\%；於波動性較高、流動性較低的 OTC 市場同樣取得 21.29\% 年化報酬、夏普比率 5.13 與 Calmar 比率 6.45，明顯超越動能與 PCA 基準，顯示集成樹模型能有效捕捉傳統線性架構難以掌握的非線性因子交互作用。本研究首次系統性驗證美股預測因子於臺灣市場的可現性，提出減少極端值與非預期風格曝險的嚴謹前處理與模型比較流程，並提供全市場與細分市場實證，證明 LightGBM 目前是臺灣多因子選股的最佳實務途徑，未來可進一步納入交易成本、槓桿限制與深度學習合成因子，縮短學術與可投資實務的距離。
This study examines cross‑sectional return predictability in the Taiwan equity market from 2010 to 2025. We analyze a curated library of 52 firm‑level predictors spanning valuation, growth, profitability, quality, technical, and liquidity dimensions. Daily data from the Taiwan Economic Journal (TEJ) are processed through a three‑step pipeline—Median Absolute Deviation (MAD) clipping, size‑ and industry‑neutralization, and Z‑score standardization—to ensure signal comparability and robustness.
Single‑factor tests show that value and quality variables deliver the highest Information Coefficients (IC) and Information Ratios (IR), whereas momentum, size, and risk‑oriented factors exhibit more volatile performance. To synthesize information across predictors, we compare four classes of aggregation techniques: (i) equal‑weight combinations, (ii) Principal Component Analysis (PCA), (iii) cross‑sectional and time‑series factor momentum (CSFM / TSFM), and (iv) three gradient‑boosting rankers—CatBoost, XGBoost, and LightGBM. Among these, LightGBM attains the strongest out‑of‑sample results, recording an annualized return of 15.90%, a Sharpe ratio of 2.50, and a maximum drawdown of only −5.90% on the whole‑market sample.
Robustness tests on the more volatile and less liquid OTC segment confirm the superiority of machine‑learning models: LightGBM still achieves a 21.29% annualized return, a Sharpe ratio of 5.13, and a Calmar ratio of 6.45, comfortably outperforming traditional momentum and PCA benchmarks. These findings underscore the adaptability of ensemble‑tree models in emerging markets and highlight their capacity to capture nonlinear factor interactions that conventional linear or momentum frameworks may overlook.
Our contributions are three‑fold: (i) we provide the first comprehensive transferability test of U.S.‑validated predictors to Taiwan, (ii) we propose a rigorous preprocessing and model‑comparison protocol that mitigates extreme values and unintended style exposures, and (iii) we furnish market‑wide and segment‑specific evidence that LightGBM currently offers the most effective route to multi‑factor stock selection in Taiwan. Future research can extend this framework by incorporating dynamic transaction‑cost models, leverage constraints, and deep‑learning‑based factor integrators to further bridge the gap between academic insight and investable practice. |
| Reference: | [1] Dhingra, V., Sharma, A., & Gupta, S. K. (2023). Sectoral portfolio optimization by judicious selection of financial ratios via PCA. Optimization and Engineering, 25(3), 1431–1468.
[2] Engelberg, J., McLean, R. D., Pontiff, J., & Ringgenberg, M. C. (2023). Do cross‑sectional predictors contain systematic information? Journal of Financial and Quantitative Analysis, 58(3), 1172–1201.
[3] Gu, S., Kelly, B. T., & Xiu, D. (2020). Empirical asset pricing via machine learning. The Review of Financial Studies, 33(5), 2223–2270.
[4] Gupta, T., & Kelly, B. T. (2018). Factor momentum everywhere (SSRN Scholarly Paper No. 3300728). Social Science Research Network.
[5] Scruggs, J. T. (2021, August). Does neutralizing style factors help or hurt? The Journal of Investing, 30(3).
[6] Wei, X., Tian, Y., Li, N., & Peng, H. (2024). Evaluating ensemble learning techniques for stock index trend prediction: A case of China. Portuguese Economic Journal, 23(3), 505–530. |
| Description: | 碩士
國立政治大學
金融學系
112352034 |
| Source URI: | http://thesis.lib.nccu.edu.tw/record/#G0112352034 |
| Data Type: | thesis |
| Appears in Collections: | [金融學系] 學位論文
|
Files in This Item:
| File |
Description |
Size | Format | |
|---|
| 203401.pdf | | 2288Kb | Adobe PDF | 0 | View/Open |
|
All items in 政大典藏 are protected by copyright, with all rights reserved.
|
