臺灣 AI 供應鏈電子股投資組合建構與期貨避險策略研究

Abstract

中文摘要 本研究以 CDIO（構想-設計-實現-運作）框架為核心，針對臺灣 AI 供應鏈電子產業，比較主觀裁量型（組合 A）與數據驅動型（組合 B）兩種投資組合之績效差異，以 R 語言模擬驗證期貨避險工具之實務效果。樣本期間為 2024 年 5月至 2026 年 5 月（484 個交易日） ，資料來源為台灣經濟新報（TEJ）資料庫，涵蓋被動元件、載板、散熱、記憶體、玻纖布及 CPO 六大 AI 供應鏈族群共 15 檔個股。在建構投資組合的部份，組合 A 採主觀選股，加權 Beta 值達 1.12（相對TWSE 大盤），呈集中進攻格局；組合 B 採 Top-Down 兩階段 Markowitz 效率前緣最佳化，加權 Beta 值為 1.14，具較佳分散效果。實證結果顯示，組合 B 累積報酬 2,064%、Sharpe Ratio 3.68，顯著優於組合 A 之 521%與 2.78，顯示在風險調整報酬上具明顯優勢。在期貨避險方面，電子期貨因與 AI 中小型股相關性極低（R²≈1.1%〜4.5%）幾乎無避險效果；個股期貨效果顯著提升，三商品組合對組合 A 的 R²達 74.7%。另外，Hull 公式在低相關情境下高估合約口數達 11.1 至 18.8 倍，OLS 最小變異避險比率為更具實務可行性之替代方法。同時，本文以類熊市壓力測試（2025 年 2 月 24 日至 4 月 14 日）進一步驗證期貨避險之保護效能：組合 A 避險後損失由-25.3%縮減至-4.5%（改善 20.8%） ，組合 B 由-28.6%縮減至-17.9%（改善 10.7%）。本研究結論指出，數據驅動型（組合 B）在績效上優於主觀裁量型（組合 A） ；個股期貨為精準避險工具，但投資組合分散程度越高，避險邊際效益相對降低，成本效益比亦需審慎評估。

Abstract This study applies the CDIO framework to empirically compare two portfolio construction approaches and evaluate futures hedging effectiveness within Taiwan's AI supply chain electronics sector. Using TEJ data from May 2024 to May 2026 (484 trading days), 15 stocks across six subsectors (passive components, IC substrate, thermal management, memory, fiberglass fabric, and CPO) are used to construct Portfolio A (discretionary) and Portfolio B (data-driven). Portfolio A concentrates in high-Beta leaders with a weighted Beta of 1.12 vs. TWSE. Portfolio B employs a Top-Down two-stage Markowitz optimization, yielding a weighted Beta of 1.14 with superior diversification. Portfolio B substantially outperformed: cumulative return 2,064% and Sharpe Ratio 3.68, versus Portfolio A's 521% and 2.78. For hedging, three instrument tiers were evaluated. Electronics futures proved ineffective (R² ≈ 1.1%–4.5%). Single-stock futures improved substantially, with a three-instrument combination achieving R² = 74.7% for Portfolio A. A key finding is that Hull's formula overestimates required contracts by 11.1× (Portfolio A) to 18.8× (Portfolio B) under low correlation, while the OLS minimum-variance hedge ratio provides a more practical alternative. A bear-market stress test (2025/02/24–04/14) confirmed hedging value: Portfolio A's loss narrowed from −25.3% to −4.5% (protecting 20.8 ppts); Portfolio B's from −28.6% to −17.9% (protecting 10.7 ppts). The study concludes that data-driven construction outperforms discretionary selection on a risk-adjusted basis, while greater portfolio diversification simultaneously reduces the marginal effectiveness of single-stock futures hedging.