Background: Adherence to tamoxifen among women diagnosed with hormone receptor positive metastatic breast cancer (mBC) can improve survival and minimize recurrence. Screening for non-adherence at treatment initiation may support personalized care, improve health outcomes, and minimize cost of care. This study aimed to use real world data (RWD) and machine learning (ML) methods to classify tamoxifen non-adherence. Methods: A cohort of women diagnosed with incident mBC from 2012 to 2018 were identified from Truven MarketScan Commercial Claims and Encounters and Medicare supplemental administrative claims databases. Patients with < 80% proportion of days coverage (PDC) in the year following treatment initiation were classified non-adherent. Training and internal validation cohorts were randomly generated (4:1 ratio). Clinical procedures, comorbidity, treatment and healthcare encounter features in the year prior to treatment initiation were used to train logistic regression, boosted logistic regression, random forest, and feed forward neural network models and internally validated based on area under receiver operating characteristic (AUROC) curve. The most predictive ML approach was evaluated to assess feature importance. Results: A total of 3,022 patients were included with 39.9% classified as non-adherent. All ML models had moderate predictive accuracy. Logistic regression (AUROC 0.64) was easily interpreted with sensitivity 94% (95% confidence interval [CI]: 0.89, 0.92) and specificity 0.31 (95% CI: 0.29, 0.33). The model accurately classified adherence (negative predictive value 88.7%) but was non-discriminate for non-adherence (positive predictive value 47.7%). Variable importance identified top predictive factors, including patient features (≥55 years old) and pre-treatment procedures (lymphatic nuclear medicine, radiation oncology, arterial surgery). Conclusions: ML using baseline administrative data predicts tamoxifen adherence. Baseline claims may not be sufficient to predict treatment non-adherence. Further validation with enriched longitudinal data may improve model performance for incorporation of predictions into clinical decision support.