Markov State Models (MSMs) are widely used to extract kinetic properties from molecular dynamic (MD) simulations of proteins. These models discretize the conformational space to describe the MD trajectory via a set of metastable states and transition rates. Conventional MSM constructions typically rest on projections of the atomic coordinates onto a suitably chosen subspace of the full configuration space. An widely used projection, considered the gold standard by many, employs the time-lagged independent correlation analysis (TICA). TICA aims at identifying slow processes by maximizing the autocorrelation along its components with respect to a previously defined lag time. We asked whether analysis of the TICA projections alone suffices to judge the quality of the resulting MSM. As a starting point to answer this question, we carried out TICAs of 2000 unbiased 1μs MD trajectories—10 trajectories each for 200 globular proteins with sequence lengths from 35 to 399 amino acids. For strikingly many trajectories the TICA projections along the slowest two components resemble cosines with half and full period and do not show pronounced metastable states. We used the cosine content—originally defined as the inner product between the projection on a PCA component with a cosine [Berk Hess, 2002]—to quantify the resemblance and analyze the relation between cosine content and protein properties. We found that, on average, the cosine content increases with protein size indicating that the cosine content can be used to detect undersampling and, hence, overfitted MSM models.