In order to alleviate congestion and make the most of public transport in smart cities, traffic and mobility predictions are very necessary. The enormous computational complexity of Long Short-Term Memory (LSTM) networks makes it difficult to employ them in real-time applications, despite the fact that these networks are particularly effective at forecasting traffic flows by identifying long-term correlations in time-series data. Through the incorporation of stochastic connectivity into recurrent layers, this study proposes a novel Randomly Connected Long Short-Term Memory (RCLSTM) model that reduces the amount of processing that is considered necessary. This design maintains the capability to capture crucial temporal correlations while also lowering the amount of learnable parameters. As a result, the model becomes more effective and suitable for real-time prediction. Using real-world traffic datasets, such as GPS traces and sensor data, we put the RCLSTM model through its paces and compared it against traditional LSTM, GRU, and ARIMA models. Because the RCLSTM model shows increased or equal accuracy while substantially reducing the amount of processing overhead, it is suitable for use in real-time pedestrian and vehicular traffic management systems in metropolitan areas.