WHY “Vector Autoregression (VAR)”? Many social (trends, polls), sport (scores), economic (indicators), business (sales), natural, and engineering events can be represented quantitatively by (hourly, daily, weekly, monthly, etc.) time-series. Researches and reports demonstrate that many of these time-series interact one another directly or through underlying common factors and many time-series are influenced by their own histories as well. VAR model is the simplest model trying to find out predictive model of linear, mutual and temporal causalities underlying these time-series.

WHY “Long Memory VAR”? The estimation engine underlying CTVARF algorithm is “long memory vector autoregressive (LMVAR)” model. The LMVAR estimates influences of longer histories of common factors and idiosyncratic components of time-series with objective data-driven constraints. Without constraints, “estimated influences of long histories” can be contaminated by various random coincidences and lack of any predictability. The data-driven constraints make CTVARF accommodate wider ranges of values of model learning parameters, and therefore further enhance the power of machine learning.

WHY “trained model feedback loop”? To make forecasts based always on the latest updated model trained by the latest available data, we use a “continuously trained” model estimated on rolling data windows moving forward to the last time stamp of the available data. What if newly available data points arrive after training ended? With trained model feedback loop, we feed both the newly available data and the previously trained model into a new model training process. This way makes model fitting and updating efficiently by only applying them to data windows containing the new data.