WHY DFM? Dynamic factor models (DFMs) can be used to analyze and forecast large set of time-series, such as measurements and indicators of national or multinational economies, prices of products or instruments constantly traded in markets, measurements and observations of natural or engineering processes, social or political trends, and scores of sports. The evolutions of these time-series are influenced by evolutions of a number of unobserved dynamic factors commonly affecting all or many of the time-series.

WHY Long-Memory? The LMDFM algorithm estimates influence of longer histories of dynamic common factors with concepts of DPCA (dynamic principal components analysis) and 2D or 1D and inverse DFT (discrete Fourier transform). Such longer-memory estimates make LMDFM accommodate wider range of values of model learning parameters. The wider ranges can further enhance the power of machine learning.

WHAT next? Many real-world large sets of time-series are nonstationary. In general, a filtering approach could be the best for analysis and forecasts on nonstationary time-series. Bayesian filters are among more adaptive filters: more powerful due to fewer restrictive conditions. A variational Bayesian filtering is the fastest one. We, i4cast LLC, is an advanced developer in variational Bayesian filtering: we listed VBfFA algorithm here on AWS. We are now working on developing and offering a long memory dynamic factor model estimated by a variational Bayesian filtering for better forecasts.