In this paper, we propose two metrics to quantify the predictability of time series before developing a time series forecasting model, namely, spectral predictability score and maximum Lyapunov exponent. Unlike traditional model evaluation metrics, these metrics assess the inherent predictability characteristics of data before attempting a forecast. The spectral predictability score evaluates the strength and regularity of frequency components of a time series, while the Lyapunov exponent quantifies the chaos and stability of the system generating the data. We evaluate the effectiveness of these metrics on synthetic and real time series from the M5 forecasting competition dataset. The results show that these two metrics accurately reflect the inherent predictability of time series and have strong correlations with the actual forecasting performance of various models.
