Outliers Z score greater than 3.

Forecasting speculative stock prices is essential for effective investment risk management and requires innovative algorithms. However, the speculative nature, volatility, and complex sequential dependencies within financial markets present inherent challenges that necessitate advanced techniques. In this regard, a novel framework, ACB-XDE (Attention-Customized BiLSTM-XGB Decision Ensemble), is proposed for predicting the daily closing price of speculative stock Bitcoin-USD (BTC-USD). The proposed ACB-XDE framework integrates the learning capabilities of a customized Bi-directional Long Short-Term Memory (BiLSTM) model with a novel attention mechanism and the XGBoost algorithm. The customized BiLSTM leverages its learning capabilities to capture complex sequential dependencies and speculative market trends. Meanwhile, the new attention mechanism dynamically assigns weights to influential features based on volatility patterns, thereby enhancing interpretability and optimizing effective cost measures and volatility forecasting. Moreover, XGBoost handles nonlinear relationships and contributes to the proposed ACB-XDE framework’s robustness. Furthermore, the error reciprocal method improves predictions by iteratively adjusting model weights based on the difference between theoretical expectations and actual errors in the individual attention-customized BiLSTM and XGBoost models. Finally, the predictions from both the XGBoost and attention-customized BiLSTM models are concatenated to create a varied prediction space, which is then fed into the ensemble regression framework to improve the generalization capabilities of the proposed ACB-XDE framework. Empirical validation of the proposed ACB-XDE framework involves its application to the volatile Bitcoin market, utilizing a dataset sourced from Yahoo Finance (Bitcoin-USD, 10/01/2014 to 01/08/2023). The proposed ACB-XDE framework outperforms state-of-the-art models with a MAPE of 0.37%, MAE of 84.40, and RMSE of 106.14. This represents improvements of approximately 27.45%, 53.32%, and 38.59% in MAPE, MAE, and RMSE respectively, over the best-performing attention-BiLSTM. The proposed ACB-XDE framework presents a technique for informed decision-making in dynamic financial landscapes and demonstrates effectiveness in handling the complexities of BTC-USD data.

对投机性金融资产价格进行预测，是做好有效投资风险管理的关键前提，同时离不开创新算法的支撑。然而，金融市场固有的投机属性、剧烈波动性以及复杂的序列依赖关系，为预测任务带来了诸多先天挑战，亟需借助高阶技术手段予以解决。为此，本文提出一种全新框架——ACB-XDE（Attention-Customized BiLSTM-XGB Decision Ensemble），用于预测投机性加密货币比特币兑美元（Bitcoin-USD，BTC-USD）的每日收盘价。所提ACB-XDE框架将定制化双向长短期记忆网络（Bi-directional Long Short-Term Memory，BiLSTM）的学习能力，与新型注意力机制以及XGBoost算法进行融合。该定制化BiLSTM可依托自身学习能力，捕捉复杂的序列依赖关系与投机性市场趋势。与此同时，该新型注意力机制可基于波动模式，对关键特征动态分配权重，进而提升模型可解释性，并优化有效成本度量与波动预测效果。此外，XGBoost能够有效处理非线性关系，助力提升ACB-XDE框架的鲁棒性。进一步地，误差倒数法可基于注意力定制化BiLSTM与XGBoost单个模型的理论预期值与实际误差之差，迭代调整模型权重，从而优化预测效果。最后，将XGBoost与注意力定制化BiLSTM模型的预测结果进行拼接，构建多样化预测空间，随后将其输入至集成回归框架中，以提升ACB-XDE框架的泛化能力。针对所提ACB-XDE框架的实证验证，将其应用于波动剧烈的比特币市场，所用数据集取自雅虎财经（Yahoo Finance）的比特币兑美元数据，时间跨度为2014年10月1日至2023年1月8日。所提ACB-XDE框架的表现优于当前最先进的模型，其平均绝对百分比误差（Mean Absolute Percentage Error，MAPE）为0.37%，平均绝对误差（Mean Absolute Error，MAE）为84.40，均方根误差（Root Mean Squared Error，RMSE）为106.14。相较于表现最优的注意力-BiLSTM模型，该框架在MAPE、MAE与RMSE指标上分别实现了约27.45%、53.32%与38.59%的性能提升。所提ACB-XDE框架为动态金融环境下的科学决策提供了可行技术方案，并验证了其在处理BTC-USD数据复杂特性方面的有效性。