Background: In prognostic research, prediction rules are generally statistically derived. However the composition and performance of these statistical models may strongly depend on the characteristics of the derivation sample. The purpose of this study was to establish consensus among clinicians and experts on key predictors for persistent shoulder pain three months after initial consultation in primary care and assess the predictive performance of a model based on clinical expertise compared to a statistically derived model. Methods. A Delphi poll involving 3 rounds of data collection was used to reach consensus among health care professionals involved in the assessment and management of shoulder pain. Results: Predictors selected by the expert panel were: symptom duration, pain catastrophizing, symptom history, fear-avoidance beliefs, coexisting neck pain, severity of shoulder disability, multisite pain, age, shoulder pain intensity and illness perceptions. When tested in a sample of 587 primary care patients consulting with shoulder pain the predictive performance of the two prognostic models based on clinical expertise were lower compared to that of a statistically derived model (Area Under the Curve, AUC, expert-based dichotomous predictors 0.656, expert-based continuous predictors 0.679 vs. 0.702 statistical model). Conclusions: The three models were different in terms of composition, but all confirmed the prognostic importance of symptom duration, baseline level of shoulder disability and multisite pain. External validation in other populations of shoulder pain patients should confirm whether statistically derived models indeed perform better compared to models based on clinical expertise. © 2011 Vergouw et al; licensee BioMed Central Ltd.