A Search-Then-Forecast Transformer Framework for Mid-Term Stock Price Prediction: An Empirical Case Study on the Chinese A-Share Market

This paper proposes a search-then-forecast framework for mid-term (20-trading-day) stock price forecasting and evaluates it on Chinese A-share data. The framework combines a multi-distance voting-based similar-stock search for sample augmentation, a sample-level PCA–ICA feature reconstruction, and a Transformer encoder–decoder whose decoder is initialised at inference with the single-step return at the end of the observation window rather than the conventional zero-padding. Using Luoyang Molybdenum (stock code 603993) from the SSE 180 pool as a single-stock case study, we compare six configurations—TransE, TransED (nhead = 3 and 6), BiLSTM, ARMAGARCH, and a TransED-Embed ablation—across ten observation window lengths, and introduce two baseline-referenced metrics, R2 hist and MASEnaive, to address the limited interpretability of standard R2 and MASE on non-stationary financial series. ARMAGARCH attains the lowest root mean squared error (RMSE) across all tested windows, outperforming the best deep learning model (TransE) by 1.4% to 17.0%; MASEnaive further reveals that most deep learning models fail to surpass a random-walk naive baseline. Observation window length and model architecture exhibit a clear interaction, and a smaller internal Transformer dimension does not hurt performance. Within this single-stock case study, the findings suggest that parsimonious statistical models can match or outperform highly parameterised deep learning architectures for mid-term Chinese A-share forecasting.