Insomnia is a prevalent sleep disorder, yet treatment response varies, highlighting the need for personalized interventions. While effective, predicting individual outcomes in therapies like telephone-guided Sleep Restriction Therapy (SRT) remains challenging. Existing models often overlook daily behavioral variability. This study examined whether adherence-related patterns, captured as variability in sleep diary data, improve the prediction of Insomnia Severity Index (ISI) outcomes. Linear (Ridge, Lasso) and non-linear (XGBoost, Random Forest) models were trained using a nested cross-validation framework accounting for both the temporal and participant-level structure of the data. Models were evaluated based on predictive accuracy. Two feature sets were compared: traditional weekly mean-based predictors and a combined set including variability-based measures. Although variability-based features did not lead to statistically significant improvements, non-linear models tended to perform better when they were included. Further feature importance analyses revealed that these predictors contributed meaningfully to model performance. These findings highlight the potential value of day-to-day variability for predicting insomnia treatment outcomes.