Panic disorder is a chronic anxiety condition with sudden and unpredictable symptoms, which makes it important to detect early signs of an upcoming attack. Digital phenotyping has become a useful way to track symptoms in real time, but most existing studies focus on classifying people rather than forecasting symptoms within a person over time. They also rarely compare their models to simple baselines, which makes it hard to know whether complex models are adding any value. 

The present study investigated whether panic symptom severity can be forecasted day by day for individual participants using time-series methods, and whether adding contextual variables improves prediction over models that rely only on past panic values. Using two years of digital phenotyping data from a publicly available dataset (N = 43, reduced to N=8 after data quality filtering), ARIMA, ARIMAX and a Naïve model were fitted for each participant under a shared preprocessing pipeline and evaluated using expanding-window one-step-ahead cross-validation. ARIMAX did not outperform ARIMA, and the Naïve model stayed competitive across both analyses. This shows how hard it is to forecast rare, episodic events like panic at a daily level.