Abstract Background Cancer Antigen 125 (CA125) is currently the best available ovarian cancer screening biomarker. However, CA125 has been limited by low sensitivity and specificity in part due to normal variation between individuals. Personal characteristics that influence CA125 could be used to improve its performance as screening biomarker. Methods We developed and validated linear and dichotomous (≥35 U/mL) circulating CA125 prediction models in postmenopausal women without ovarian cancer who participated in one of five large population-based studies: Prostate, Lung, Colorectal, and Ovarian Cancer Screening Trial (PLCO, n = 26,981), European Prospective Investigation into Cancer and Nutrition (EPIC, n = 861), the Nurses’ Health Studies (NHS/NHSII, n = 81), and the New England Case Control Study (NEC, n = 923). The prediction models were developed using stepwise regression in PLCO and validated in EPIC, NHS/NHSII and NEC. Result The linear CA125 prediction model, which included age, race, body mass index (BMI), smoking status and duration, parity, hysterectomy, age at menopause, and duration of hormone therapy (HT), explained 5% of the total variance of CA125. The correlation between measured and predicted CA125 was comparable in PLCO testing dataset (r = 0.18) and external validation datasets (r = 0.14). The dichotomous CA125 prediction model included age, race, BMI, smoking status and duration, hysterectomy, time since menopause, and duration of HT with AUC of 0.64 in PLCO and 0.80 in validation dataset. Conclusions The linear prediction model explained a small portion of the total variability of CA125, suggesting the need to identify novel predictors of CA125. The dichotomous prediction model showed moderate discriminatory performance which validated well in independent dataset. Our dichotomous model could be valuable in identifying healthy women who may have elevated CA125 levels, which may contribute to reducing false positive tests using CA125 as screening biomarker.