Modelling popularity for academic fields have been an ongoing study. This is especially important for emerging fields as accurate models allow efforts and resources to be efficiently utilised on promising research directions. Existing modelling methods mainly face at least one of the following three challenges: Using domain specific binary classifications on topics to be ether emerging or non-emerging leading to low generalizability. Having a biased and restricted scope of investigation due to topic terms requiring manual mining from a limited number of documents. Neglecting the effect of "cold start" when utilising a field’s historical features as inputs for popularity forecasting especially when the field is emerging and possesses limited historical data. In this paper, we build upon existing work to propose a forecasting algorithm addressing all three challenges. Firstly, we define popularity forecasting as a multivariate regression problem. Next, by combining the utilisation of existing academic databases, time specific node embeddings, and dynamic time warping, we extract concurrently trending neighbour fields whose trending pattern are similar to the field of interest. Lastly, multivariate forecasting is conducted using long short-term memory (LSTM) and dual attention recurrent neural networks (DA-RNN). Experimental results on 10 emerging and non-emerging fields of study showcases the existence and various dynamics of "cold start". Additionally, the proposed algorithm is also shown to greatly reduce RMSE, MAE, and MAPE against traditional methods for emerging fields while retaining similar performance for non-emerging fields. This validates the significance of these challenges against existing methods and provides insight on the dependency structure of emerging topics with their historical features.