傳統的股市預測模型多半致力於預測絕對價格，然在充滿雜訊的金融市場中，此 類模型極易陷入滯後效應，在實務交易中缺乏預警能力。因此，將預測標的轉向捕捉 具備實質獲利空間之「波段反轉點（Turning Points）」，已成為量化交易之核心趨 勢。然而，實務導向的嚴格標註條件必然導致正負樣本極度不平衡（Imbalanced Data）；此外，純技術面模型在面臨市場恐慌下殺時，容易產生過早進場的「接刀 （Catching a falling knife）」風險。 為解決上述痛點，本研究以台灣權值股台積電（2330.TW）為實證對象，提出一套 結合一維卷積神經網路（1D-CNN）與SBERT文本語義萃取之「多模態深度學習預測系 統」。研究方法首先利用Focal Loss焦點損失函數取代傳統交叉熵，透過動態權重分 配，解決正樣本佔比僅7%（正負樣本計178筆對2365筆，比例約13:1）之嚴重類別 不平衡問題；其次，將非結構化之每日科技新聞轉化為768維語義向量，並透過主成 分分析（PCA）壓縮至32維後，與量價動能進行特徵級融合（Feature-level Fusion）。在模型決策端，本研究屏棄靜態閾值，改採前5%之動態百分位閾值，並疊加 VIX 恐慌突波與RSI超賣（RSI<35）作為雙重防禦濾網，同時引入5日訊號冷卻期。 此外，本研究另透過消融實驗驗證ATR動態波動門檻與VIX／RSI雙重恐慌濾網之 獨立貢獻，結果顯示動態百分位閾值本身已是精準率提升之主要來源，恐慌濾網則在 此基礎上進一步強化篩選品質，相關發現詳見第四章之消融分析。 實證結果顯示，在允許1日誤差之實務容錯評估下，純量價模型之容錯精準率為 0.5；而導入新聞語義特徵之多模態模型，其容錯精準率顯著提升至0.75，精準率絕對 提升幅度達32個百分點（相對提升50%）。研究結論證實，非結構化之新聞文本能有效 反映市場悲觀情緒之延續性，在神經網絡中發揮關鍵的「煞車機制」，成功抑制技術指 標超跌所引發之誤判。本研究提出之多模態預警框架，不僅有效降低左側交易之破底 風險，更大幅提升了捕捉底部反轉點之實務勝率。;Traditional stock market prediction models typically focus on forecasting absolute prices. However， in noisy financial markets， such models often fall into lagging effects and lack early warning capabilities in practical trading. Therefore， shifting the prediction target to capturing "turning points" with substantial profit margins has become a core trend in quantitative trading. Nevertheless， strict profit-oriented labeling inevitably leads to highly imbalanced data. Furthermore， price-only technical models are prone to the risk of "catching a falling knife" (premature entry) during market panic sell-offs. To address these challenges， this study proposes a "multimodal deep learning prediction system" combining 1-Dimensional Convolutional Neural Networks (1D-CNN) and SBERT semantic extraction， using Taiwan′s leading stock， TSMC (2330.TW)， as the empirical target. Methodologically， the Focal Loss function is employed instead of traditional cross-entropy to overcome the severe class imbalance， where positive samples (reversal points) account for only 7% of the dataset (178 positive vs. 2，365 negative instances， approximately 13:1 ratio). Additionally， unstructured daily tech news is transformed into 768-dimensional semantic embeddings， which are compressed to 32 dimensions via Principal Component Analysis (PCA) before being integrated with price-volume dynamics through feature-level fusion. At the decision-making stage， instead of static thresholds， this study adopts a dynamic top 5% percentile threshold， overlaid with VIX spikes and RSI oversold conditions (RSI < 35) as dual defense filters， supplemented by a 5-day signal cooldown period. The empirical results demonstrate that under a practical 1-day tolerance evaluation metric， the price-only model achieves a tolerant precision of 0.5. In contrast， the multimodal model incorporating news semantic features significantly improves the tolerant precision to 0.75， representing an absolute improvement of 25 percentage points (relative improvement of 50%). The findings confirm that unstructured news texts effectively reflect the continuity of market pessimism， acting as a crucial "brake mechanism" within the neural network to suppress misjudgments triggered by oversold technical indicators. The proposed multimodal early warning framework not only effectively mitigates the downside risks of left-side trading but also substantially enhances the practical win rate of capturing bottom turning points.
