Background Many hospitals use predictive scores to identify a person’s risk of inpatient falls, pressure injury and malnutrition despite evidence of limited predictive accuracy. Aim To examine whether we could improve predictive accuracy by generating a score combining all components of currently‐used tools. Methods We performed a retrospective, cross‐validation study in a single sub‐acute (geriatrics and rehabilitation) hospital extracting data regarding hospital risk scores, and incidence of falls, pressure injury, and malnutrition from January 2014 to June 2016. The sample was randomly halved into training and testing datasets. For each harm outcome, model fit was examined using Area Under Receiver Operating Characteristic Curves (AUC) and proportions of people reclassified based on a combined score were calculated. Secondary analyses explored the predictive performance of individual question‐responses. Results Data was available for 4487 admissions (median age 83.0 years). A total of 667 (15%) people had at least one fall, 499 (11%) had at least one pressure injury and 20 (0.4%) malnutrition. The currently‐used tools had, at best, moderate ability to predict risk of harm outcomes (AUC 0.56‐0.73). Testing of the combined score models resulted in minimal change in AUC (<5.1%) and did not add value to risk category reclassification. Most of the predictive ability of the currently‐used tools relied on the performance of two individual question‐responses. Conclusions Combining scores or reducing to two item question‐responses did little to change predictive accuracy. This study highlights the limitations of hospital harm predictive scores and emphasises the importance of rigorous testing of predictive scores.