Parametrik olmayan MİDAS regresyon

Nonparametric MIDAS (Mixed Data Sampling) regression methods, which allow the analysis of high-frequency explanatory variables, are examined. As a frontier study of the nonparametric MIDAS regression method, the thesis presents a pioneering effort with big and stream data. Lit-review of parametric and nonparametric MIDAS regression, a performance comparison via the designed experiment and findings of real-case implementation are presented. The drawbacks of the widely adopted temporal aggregation and iteration methods in the conversion of high-frequency variables, and relevant proposals involving MIDAS regression are presented. Also, Bridge Equations and State Space Models are examined. Then, autoregressive time series, MIDAS parametric weighting functions, data matrices and MIDAS types are discussed; Almon Polynomial, Exponential Almon, Beta and Stepwise weighting functions are explained. Next, spline and kernel regression models, MIDAS smoothing least square (SLS) are presented. The potential usability of kernel regression methods with MIDAS is discussed. In the simulation stress tests with negative-positive, short-long lag lengths, and variance scaling was performed, and results were compared. Random data generation via designed weighting functions for stress testing was used. Finally, MIDAS regression methods were run on the daily COVID-19 cases recorded in Turkey as the dependent variable and the number of hourly shared Turkish Twitter messages with COVID-19 content as the high-frequency independent variable. Accordingly, nonparametric MIDAS is more effective than parametric methods in cases with more impact in the long term compared to short-term, negative-positive values, and long lag times. Limitations of kernel regression together with computation burden should be addressed in future studies.