Accurate prediction of protein secondary structure is essential for reliable tertiary structure prediction and peptide design. To address this, multiple algorithms have been proposed. The conformation of a residue within a polypeptide chain is strongly influenced by its immediate neighbors. The conformational tendencies of 20 residues in the presence of their first neighbors, constituting a total of 8000 tripeptides, were calculated. Using the propensity values of 8000 tripeptide variants, we propose an accurate method for secondary structure prediction. Various machine learning (ML) models were built using propensities, position-specific scoring matrices (PSSM) and amino acid binary features to predict three-state (Q3) and eight-state (Q8) secondary structures. The results suggest that the XGBoost ML model produced highly accurate predictions, achieving accuracies of 93% and 88% for Q3 and Q8 state, respectively for validation dataset CB513. A program for prediction called S8kPred is developed. This utility along with consensus secondary structure prediction is available online at https://www.s8kpred.in.