Commodity futures prices exhibit high nonlinearity, non-stationarity, and volatility. The sentiment polarity extracted from news texts has significantly improved the timeliness and accuracy of forecasting. However, massive news texts contain a significant amount of noise information, which severely impacts the stability of forecasting. Therefore, sentiment analysis is conducted on news texts related to futures prices, and a weighted double-well potential surface support vector regression (WDWPS-SVR) model is proposed to improve prediction robustness for commodity futures prices. The model generates a quartic double well potential surface to conduct regression on the data in the original feature space, capturing the complex data characteristics in high-frequency data. Besides, the weight of each training sample are calculated using a weight function to assess their importance in the fitting process, reducing the interference of news texts noise in parameter estimation. The model is applied to predict soybean futures closing prices on the Dalian Commodity Exchange from 2019 to 2022. The experimental results show that: 1) Sentiment variables extracted from news texts effectively explain high-frequency price volatility, improving prediction timeliness; 2) The WDWPS-SVR model’s strong nonlinear learning ability enhances prediction accuracy by fitting fine-grained price fluctuations; 3) A weighting approach reduces noise from news texts, boosting model robustness and prediction stability.