Real-world financial time series often contain both linear and nonlinear patterns. However, traditional time series analysis models, such as ARIMA, hold the assumption that a linear correlation exists among time series values while leaving nonlinear relation into error terms. Based on financial theories, we argue that investor sentiment is the main contributor to nonlinear pattern of stock time series. Furthermore, we propose a sentiment-based hybrid model (SLNM) to better capture nonlinear information in stock time series. According to the forecasting experimental results, SLNM exhibits the sensitivity to sentiment environments, which in turn supports the argument that investor sentiment is the main source of nonlinear pattern in stock time series. For those stocks that are in top 10 of CAR Ranking List ─ these stocks are more likely pursed by emotional investors and thus in optimistic sentiment environment, SLNM improves forecasting performance dramatically: Increase Direction Accuracy by 40% and reduce RMSE by 19.3%. While, for those that are in bottom 10 of CAR Ranking List─ these stocks defer more emotional investors from further participating in stock trading and thus in pessimistic sentiment environment, SLNM has a fair improvement on performance: Hold the similar Direction Accuracy and reduce RMSE only by 2.5%. All these indicate that investor sentiment play a key role in stock return forecasting. Our work sheds light on the research of sentiment-based prediction models.