Skip to main content Link Search Menu Expand Document (external link)

Sensei

v0.2.4

Sensei is a data-driven top-down modeling engine that provides projections, as well as parameter and uncertainty estimates for time series where variable relationships are defined by a CAG structure. Sensei models the multivariate time series as a set of univariate time series models with “regressors” defined by the CAG structure. That is to say, each node in the CAG gets its own model, which consists of a univariate time series model which parameterizes the trend and seasonality of the time series, as well as coefficients on the incoming edges from one-step-behind values. We compute parameters and projections step-wise, i.e. every model estimates one-time-step-ahead projections iteratively.

The model we use for univariate time series is the Damped Local Trend (DLT) model , variants of which are generally top performers on the M3 competition. The parameters and uncertainty are estimated using Maximim A Posteriori (MAP) which offers good accuracy at interactive speed.