TIM's architecture tries to follow all the specifics of time series modelling to give you the best possible accuracy in the smallest amount of time. To learn about the motivation for its design, refer to Time Series Forecasting to find out more about why and how time series are different from other data.
TIM acts like a general time series model building platform, but given our experience in specific industries like load forecasting, we can share our knowledge from these use cases to tell you which data are best plugged alongside your target variable to achieve the highest accuracy.
TIM is designed to chew all the predictors you could possibly plug into it and understand which are important and which are useless. This is natural, because TIM itself generates many new predictors (features) from the original set. These, too, might be useful or useless at the first glance. We call this process Expansion. To achieve stability of our models and to fight overfitting, TIM reduces this newly created set of predictors to just a small subset. This we call Reduction.
To address different complexity and feature usage in different situations, TIM creates not one model, but many of them under the one Model Zoo. To ensure, that this is computationally efficient, we have created a layer that exploits similarities in different models and their situations to avoid redundant computations.
In production, the resulting Model Zoo is then prepared to dispatch the most appropriate model without you even noticing.
In anomaly detection, TIM is firstly used to forecast what should be happening and anomaly detection layer can then assess deviation from normality.