Motifs in multivariate time series reveal critical non-periodic behaviors in biophysiological, geophysical, urban, and societal systems. This work takes motif analysis one step further by exploring the recurrence dynamics of multidimensional motifs to enhance the forecasting of events of interest, focusing on regressing the timing of a motif’s next occurrence. This task is challenged by the inherent stochasticity of real-world system behaviors and heterogeneity of data inputs, combining the raw multivariate time series and available motif information. To address these challenges, two major hypotheses are established: i) that some irregular behaviors show indeed a form of less trivial temporal regularity, possibly described by a non-linear function; and ii) the occurrence of motifs in some systems can be anticipated by precursor signals, such as emerging traffic behaviors prior to congestion or subtle physiological patterns preceding health events. A novel methodology is proposed to expressively encode motif information and subsequently combine state-of-the-art neural processing principles to answer the target forecasting task. Experimental results demonstrate that it is feasible to accurately estimate the next occurrence of a given motif in arbitrarily complex tasks by leveraging the value embedded in the two proposed hypotheses. Furthermore, we show that augmenting the multivariate input series with motif-aware masking significantly enhances the predictive accuracy of recurrent and convolutional forecasters.