Interval-valued prediction of time series based on fuzzy cognitive maps and granular computing

Time series have yielded impressive results in numerical prediction, yet the presence of noise can significantly affect accuracy. Although interval prediction can minimize noise interference, most methods only predict upper and lower limits separately, resulting in uninterpretable predictions. In this paper, we propose a novel modeling approach for time-series interval prediction that integrates granular computing and fuzzy cognitive maps (FCMs). Granular computing transforms traditional numerical time series into interval time series. Rather than predicting interval values independently, our method mines the fuzzy relationship between information granules to obtain the affiliation matrix. During the prediction stage, an FCM-based model is established to predict the affiliation matrix. We conducted experiments on six publicly available datasets, and results demonstrate that our method reduces the impact of noise while offering improved interpretability for prediction outcomes. More importantly, our approach yields significantly lower interval prediction errors when compared to other advanced methods.