<p>Agricultural productivity and overall availability of water resources in the Indian Subcontinent are greatly influenced by the sub-seasonal variability in rainfall observed during the summer monsoon. Consequently, there is a growing demand for weather forecast products capable of predicting hydro-meteorological patterns on sub-seasonal to seasonal (S2S) timescales, particularly within a two- to six-week lead time. To address this growing demand, weather prediction models are continuously evolving, integrating advancements in data assimilation and atmospheric dynamics. However, operational model accuracy still requires improvement. In this context, studies have shown that machine learning techniques could improve forecasting accuracy and reduce uncertainties. In this work, we investigated the capability of an Extreme Learning Machine (ELM) technique to predict rainfall states up to three to four weeks ahead for the core monsoon zone of India. In comparison to other machine learning models like ANN, LSTM-CNN and linear regression model(s), the ELM technique demonstrates better performance in terms of predictability when the dataset was augmented with a special target code called the Hadamard code. The simplicity and generalization capability of the ELM algorithm enabled us to generate a large number of ensembles for a reliable tercile probabilistic forecast We have devised a novel probabilistic verification score called Realised Impact Score for evaluating the model performance especially for prolonged wet and dry events and shown that ELM model augment with Hadamard codes have performed significant well as compared to ELM model without Hadamard code for predicting anomalous Indian summer monsoon rainfall patterns.</p>