Performance statistics for the LSTM-based model.

Anchoring is widely considered one of the most robust and consistently observed effects in experimental psychology. This study employs the highest and lowest indices of the Shanghai Stock Exchange (SSE) alongside the highest and lowest bullish sentiments over a 52-week period as anchors, in conjunction with Fibonacci retracement levels, to develop a dual market–sentiment anchoring multivariate feature matrix. Based on this feature matrix, we propose a forecasting model called Market Sentiment Dual Anchoring CNN2D-ABiLSTM (MSD-CNN2D-ABiLSTM). This model employs CNN2D to extract spatial features from market and sentiment data, utilizes BiLSTM networks to process and integrate temporal features, and incorporates an attention mechanism to emphasize essential spatial and temporal information. Experimental results indicate that this model achieves a prediction accuracy exceeding 90% and an R2 value greater than 95% for lags of 1–2 trading days, enabling precise forecasting of the SSE index. Additionally, the model demonstrates effective forecasting performance for up to 10 trading days ahead, significantly outperforming traditional baseline models. Furthermore, structural sensitivity tests reveal that the extraction of local spatial features by CNN2D provides a predictive advantage over the short-term temporal features captured by CNN1D in complex market structures.锚定效应（Anchoring）被广泛认为是实验心理学中最稳健且被反复观测到的经典效应之一。本研究以52周周期内上海证券交易所（SSE）的最高、最低指数，以及同期看涨情绪的极值作为锚点，结合斐波那契回撤位（Fibonacci retracement levels），构建了双市场-情绪锚定多变量特征矩阵。基于该特征矩阵，本研究提出一种名为市场情绪双锚定二维卷积神经网络-双向注意力长短期记忆网络（Market Sentiment Dual Anchoring CNN2D-ABiLSTM，简称MSD-CNN2D-ABiLSTM）的预测模型。该模型采用二维卷积神经网络（CNN2D）提取市场与情绪数据的空间特征，借助双向长短期记忆网络（BiLSTM）处理并整合时序特征，并引入注意力机制以突出关键的空间与时序信息。实验结果表明，在1至2个交易日的滞后阶数下，该模型的预测准确率超过90%，决定系数（R²）大于95%，可实现上海证券交易所指数的精准预测。此外，该模型在最长10个交易日的预测周期内仍展现出优异的预测性能，显著优于传统基准模型。进一步的结构敏感性测试显示，在复杂市场结构中，二维卷积神经网络（CNN2D）对局部空间特征的提取，相较于一维卷积神经网络（CNN1D）捕捉的短时时序特征，具备更优的预测性能优势。