股票趨勢預測研究：結合TCAN模型與競爭式參數遷移策略

Abstract

金融市場中的股票時間序列資料具有高度非線性與動態變化特性，使模型在跨期間應用時容易受到市場狀態變化影響，進而產生預測表現不穩定之情況。基於此，本研究建立一套結合多子模型架構、跨期參數延續方法與集成決策機制之股票趨勢預測與交易模擬架構。其中，模型更新階段採用 Error-Minimization Transfer（EMT）方法，決策階段則採用多模型集成式選擇機制（Ensemble-Based Selection, EBS），並結合 ATR-based stop-loss 風險控管機制。 本研究採用 Temporal Convolutional Attention Network（TCAN）作為基礎子模型架構，並於每一個 rolling period 中透過不同隨機種子訓練多個子模型，以形成模型池。EMT 方法透過指數加權平均（Exponentially Weighted Average, EWA）機制整合子模型之歷史表現與當期排名，選取部分具有較高延續優先性之子模型進行參數延續，並與重新初始化之子模型共同構成下一期模型池。EBS 機制則整合多個子模型之預測結果，以產生最終趨勢預測結果，並進一步形成交易訊號。 在實證分析上，本研究以道瓊工業平均指數（Dow Jones Industrial Average, DJIA）之 29 檔成分股為研究標的，資料期間涵蓋 2016 年至 2023 年，並採用多期 rolling-window 架構進行評估。實驗結果顯示，EMT 方法與 Retraining 架構在趨勢分類指標上的差異相對有限，但 EMT 方法於多數交易績效指標上呈現較佳表現。進一步比較 EBS 與 SMS 兩種決策機制後，EBS 機制於本研究設定下具有較高之年化報酬、較低之最大回撤與較佳之風險調整後績效。此外，模型延續比例與 ATR 停損參數之實驗結果亦顯示，不同參數設定會影響整體交易績效，其中模型延續比例 50% 與停損參數 λ = 5.5 在本研究所比較設定中呈現較佳結果。 本研究所建立之自動化交易模擬系統整合 EMT 跨期參數延續方法、EBS 多模型集成式選擇機制與 ATR-based stop-loss 風險控管機制。實驗結果顯示，在本研究之資料範圍與實驗設定下，部分延續前一期模型參數、整合多個子模型之預測結果，並搭配適當之停損參數設定，可於所比較方法中呈現較佳之報酬與風險調整後績效。

Financial stock time series are characterized by highly nonlinear and dynamic behavior, which may cause forecasting models to be affected by changing market conditions and lead to unstable predictive performance across different periods. To address this issue, this study develops a stock trend forecasting and trading simulation framework that integrates a multi-sub-model architecture, a cross-period parameter transfer method, and an ensemble decision mechanism. Specifically, the Error-Minimization Transfer (EMT) method is employed for model updating, the Ensemble-Based Selection (EBS) mechanism is adopted for decision making, and an ATR-based stop-loss mechanism is incorporated for risk management. This study adopts the Temporal Convolutional Attention Network (TCAN) as the base sub-model architecture. In each rolling period, multiple sub-models are trained using different random seeds to construct a model pool. The EMT method employs an Exponentially Weighted Average (EWA) mechanism to integrate the historical performance and current-period rankings of the sub-models, selects a subset of sub-models with higher transfer priority for parameter transfer, and combines them with newly initialized sub-models to construct the model pool for the next period. The EBS mechanism integrates the prediction results of multiple sub-models to generate the final stock trend prediction, which is further converted into trading signals. For empirical evaluation, this study uses 29 constituent stocks of the Dow Jones Industrial Average (DJIA) from 2016 to 2023 and conducts multi-period evaluation under a rolling-window framework. The experimental results show that the differences between the EMT method and the Retraining framework in trend classification metrics are relatively limited, while the EMT method achieves better performance on most trading performance metrics. A further comparison between the EBS and SMS decision mechanisms shows that, under the experimental settings of this study, EBS achieves a higher annualized return, a lower maximum drawdown, and better risk-adjusted performance. In addition, the experiments on model transfer ratios and ATR stop-loss parameters indicate that different parameter settings affect overall trading performance. Among the settings examined in this study, a model transfer ratio of 50% and a stop-loss parameter of λ = 5.5 yield relatively better results. The automated trading simulation system developed in this study integrates the EMT cross-period parameter transfer method, the EBS ensemble decision mechanism, and the ATR-based stop-loss risk management mechanism. The experimental results show that, within the data scope and experimental settings of this study, partially transferring model parameters from the previous period, integrating the predictions of multiple sub-models, and adopting an appropriate stop-loss parameter can achieve relatively better return and risk-adjusted performance among the compared methods.