Innovations in recurrent neural networks (RNNs) for time-series modelling has enabled effective prediction of sequence data prevalent in real-world dynamic processes. For example, problem domains such as speech and language modelling, weather prediction, financial forecasting and patient healthcare monitoring have all significantly benefited from developments in RNNs over the last 20 years. Today's vast availability of data has enabled models to learn from higher quality historical information, identify loopholes, and better understand and contextualise evolving information in real-time. This is arguably extremely important as access to fast reliable predictive information empowers decision makers to prepare for and effectively respond to future opportunities or crises. Also, the simpler the forecasting model, the more amenable it is to residing on multiple mobile platforms improving accessibility of predictive power. Given this, we comparatively assess the importance of the Multi-recurrent Neural Network (MRN) with a single hidden layer against current state-of-the-art models using a range of real-world problem domains from oil price prediction to predicting the spread and mortality rate of COVID-19. We find strong evidence that the simple and shallow MRN consistently offers superior performance over the Long Short-term Memory model (LSTM), the current state-of-the-art RNN and Support Vector Machines (SVM), a more traditional statistical approach. The MRN required much fewer adjustable parameters than the LSTM to learn the task and generalise competitively. This suggests that the simpler architecture offers significant value with respect to computational resources and effective predictive abilities for real-world applications which is particularly useful given the current ubiquitous shift towards the use of Internet of Things devices.