A2030
Title: Exogenous-price adaptive routing for financial return forecasting
Authors: Jiahao Zhou - University of Macau (China) [presenting]
Abstract: Short-horizon stock return forecasting remains difficult because prices respond to both market dynamics and external information shocks. An exogenous-price adaptive routing framework is proposed for daily stock return forecasting. The model combines adjusted OHLCV data with external signals, including FinBERT-based news sentiment and macro-policy uncertainty indicators. To reduce look-ahead bias, daily external variables are lagged by one trading day, while monthly variables are aligned with a one-month availability lag. Instead of directly adding these variables to the prediction features, the model uses external signals and price-state features to generate routing weights over multiple patch scales in a Pathformer-style structure. The hidden sequence is decoded by a GRU readout to predict next-day log returns. Across Pathformer-style benchmarks and routing ablations, models are selected using validation performance before test evaluation. The results show that the proposed framework maintains competitive log-return forecasting accuracy while improving risk-sensitive performance, with higher cumulative equity and lower maximum drawdown than Pathformer-style baselines. The findings suggest that exogenous-price adaptive routing is useful when forecasting accuracy and downside risk are considered together.
