This project creates and selects predictive models for a publicly available dataset of property and casualty insurance data concerning agency performance. Written premium under a logarithmic transform was selected as the target variable, and the other 48 columns in the raw dataset were considered as potential predictors, with careful consideration of leakage — the scenario is predicting written premium for the upcoming year. Data cleaning and predictor analysis were performed to select a reasonable initial set of predictors. The models considered were OLS (Ordinary Least Squares), RF (Random Forests), and LMM (Linear Mixed Models). While an RF was selected as the leading predictive model, OLS served a useful role in rounding out the feature analysis phase of the project, and LMM quantified variance attributable to agency and parent agency identity which could not be measured using the other models. The leading RF model used the predictors identified to be valuable via predictor analysis and OLS and was tuned on a small hyperparameter grid. The RF had a test R2 of 0.8839, substantially better than the leading OLS model’s test R2 of 0.7552 and the leading LMM model’s test R2 of 0.7607. Further, the LMM revealed that agency identity accounts for 24.22% of the remaining variance after controlling for the same fixed-effect predictors used in the selected OLS and RF models. Motivated by the LMM’s result, k-means clustering was performed on the agencies, and an agency cluster predictor was added to the RF, but this showed no improvement. In conclusion, the overall leading model was an RF model tuned for selected hyperparameters using 10 selected predictors.