<p>Quality of textile yarns are decisively influenced by the fibre properties and process parameters. Various mathematical and statistical models have been used in the past for modelling the yarn properties. Machine learning (ML) models have started to make their way in the textile industry since they are able to counter several limitations of the mathematical and statistical models. In this work, an attempt has been made to predict three properties of cotton yarns (tenacity, unevenness and hairiness) using two decision tree-based ML models, namely Random Forest and XGBoost. These models yielded better results compared to traditional models and existing ML models (ANN, SVM and KNN). Particularly, XGBoost performed better than all other models and also gave the best predictions (<i>MAPE</i> = 2.53, 2.69 and 2.13; <i>R</i>^<i>2</i> = 0.776, 0.881 and 0.937; <i>RMSE</i> = 0.496, 0.731 and 0.144; and <i>MAE</i> = 0.375, 0.548 and 0.115 for tenacity, unevenness and hairiness, respectively) for all three yarn quality parameters for unseen test data. Synthetic data was also generated to improve the generalisability of the XGBoost model and it was found that the prediction accuracy improved with the synthetic data. Significance test of input features was conducted to analyse the importance of cotton fibre properties. The results reveal that there is a good agreement between the outcome of ML models and established perception of fibre-yarn property relationships.</p>