本文研究機器學習交易策略在加入無資訊雜訊特徵後，是否會先於報酬明顯惡化之 前，於交易階段呈現脆弱性。傳統樣本外驗證多著重 Sharpe 比率、平均報酬或排序式預測準確度；然而，模型可能在預測績效尚未明顯崩壞前，已透過持股排序不穩、投組換手率過高與交易成本上升等方式，對無關變數產生反應。為衡量此交易層級敏感度，本文建構雜訊換手放大指標，定義為投組換手率對加入無資訊雜訊特徵數量的斜率。 實證分析使用 TEJ 台灣上市櫃公司 2005 至 2025 年財務報表與市場資料。本文將原始產業標籤彙整為五個具經濟解釋性的產業群，以兼顧財報可比性、獲利來源、風險暴露與樣本數穩定性。對每一產業與樣本外區段，本文比較基準機器學習投組與控制式雜訊注入後的投組，並檢驗估計出的雜訊換手放大指標是否能預測後續 Sharpe 比率衰退與交易不穩定。 診斷結果顯示，五個產業中有四個產業支持此指標的診斷效度，表示雜訊換手放大指標能在多數產業捕捉具經濟意義的交易脆弱性。然而，若將該指標直接用於硬性模型篩選、門檻選擇或模型切換，樣本外績效並未穩定改善。此負結果是本文解讀的核心：該指標包含風險資訊，但不應被視為直接 alpha 訊號或通用模型選擇器。 因此，本文進一步將該指標作為漸進式交易控制的狀態變數。在 30 基點成本假設下，換手節流與換手節流結合不交易區間可在 Sharpe 比率未出現實質惡化的情況下，降低多個風險面向與交易成本。曝險縮放亦能降低下行與尾端風險，但其經濟意涵較接近風險管理 overlay。整體而言，本文貢獻在於提出機器學習交易策略雜訊脆弱性的診斷與交易控制架構，而非提出新的報酬因子。;This thesis examines whether machine-learning trading strategies become fragile at the trading stage after uninformative noise features are introduced. Conventional out-of-sample validation usually emphasizes Sharpe ratios， average returns， or ranking-based predictive accuracy. A model may nevertheless react to irrelevant variables through unstable holdings rankings， excessive portfolio turnover， and higher trading costs before its predictive performance visibly collapses. To measure this trading-level sensitivity， I construct a noise-turnover amplification measure， defined as the slope of portfolio turnover with respect to the number of injected uninformative features.
The empirical analysis uses TEJ financial statement and market data for Taiwan listed and over-the-counter firms from 2005 to 2025. Original industry labels are aggregated into five economically interpretable groups in order to balance financial-statement comparability， sources of profitability， risk exposure， and sample-size stability. For each industry and out-of-sample fold， I compare baseline machine-learning portfolios with portfolios trained under controlled noise injection， and then examine whether the estimated noise-turnover amplification measure predicts subsequent Sharpe deterioration and trading instability.
The diagnostic evidence is supported in four out of five industries， indicating that the measure captures economically meaningful trading fragility in most industry groups. However， hard model-selection， threshold-selection， and switching rules based directly on the measure do not deliver stable out-of-sample improvements. This negative result is central to the interpretation of the measure: it contains risk information， but it should not be treated as a direct alpha signal or as a universal model selector.
I therefore evaluate the measure as a state variable for gradual trading control. Under the main 30-basis-point transaction-cost assumption， turnover throttling and turnover throttling combined with a no-trade band reduce multiple risk dimensions and trading costs without a material deterioration in Sharpe ratios. Exposure scaling also lowers downside and tail risks， but its economic interpretation is closer to a risk-management overlay. Overall， the contribution of this thesis is to provide a diagnostic and trading-control framework for noise fragility in machine-learning strategies， rather than to propose a new return factor.
