This talk will present an econometrics time-series regression model for end-of-day cryptocurrency investor sentiment. We will start with what we believe constitutes challenges in Natural Language Processing for statistics, econometrics and social sciences research, before presenting our statistical framework for text-based time-series and sentiment signals construction. We will then proceed to demonstrate how we can jointly model text and finance data sampled at multiple timescales, while accounting for stylised features - in particular, long memory. In doing so, we will present our model which comprises Autoregressive Distributed-lag and Mixed Data Sampling (MIDAS) components. We will show a novel application of the Koyck transform (Koyck L. M., 1954) in a MIDAS setting, which provides tractability when the number of distributed lags tends to infinity. Finally, we will show how we resort to instrumental variable regression to successfully address the estimation issues arising after the application of the MIDAS-Koyck transform.