The purpose of this study was to develop and test personalized predictions for functional recovery after Total Knee Arthroplasty (TKA) surgery, using a novel neighbors-based prediction approach. We used data from 397 patients with TKA to develop the prediction methodology and then tested the predictions in a temporally distinct sample of 202 patients. The Timed Up and Go (TUG) Test was used to assess physical function. Neighbors-based predictions were generated by estimating an index patient’s prognosis from the observed recovery data of previous similar patients (a.k.a., the index patient’s “matches”). Matches were determined by an adaptation of predictive mean matching. Matching characteristics included preoperative TUG time, age, sex and Body Mass Index. The optimal number of matches was determined to be m = 35, based on low bias (−0.005 standard deviations), accurate coverage (50% of the realized observations within the 50% prediction interval), and acceptable precision (the average width of the 50% prediction interval was 2.33s). Predictions were well-calibrated in out-of-sample testing. These predictions have the potential to inform care decisions both prior to and following TKA surgery.