TIM is specifically designed to tackle time-series data. Therefore, extra care has been taken to assure TIM is resistant to the many pitfalls of time series modeling. Some of these pitfalls are discussed below, to illustrate how TIM succeeds in overcoming them.
Changing data patterns
Phenomena represented by time series are dynamic. Consequently, a working model is not guaranteed to stay up to date. Numerous examples in different industries illustrate this reality: changing portfolios of assets in finance, changing influential factors in trading, where changes are so frequent it is mandatory to rebuild models from scratch, changing portfolios of production assets in the utility sector…
Previously valuable suddenly become useless in these new situations. This forces users to repeat the model building process, by for example performing a new AutoML search, though this is often not an optimal solution. These new situations tend to require the identification of new significant features rather than a different modelling technique or slightly adjusted hyperparameters.
TIM empowers users to adapt to new situations by allowing models to be rebuilt or recalibrated continuously. Whereas model recalibration only adjusts the model’s parameters and leaves the model’s structure (features) intact, model rebuilding starts by identifying new features and then builds a completely new model. Through the identification of new features, the process of model rebuilding is made robust to change.
Data availability introduces another complexity, as the data availability situation at the time the model is used always needs to be considered. This situation can vary from time to time. Imagine your model uses lagged features of the target variable from 6 hours ago, i.e. y(t-6). It is possible that this data is not yet available when a forecast needs to be made. As a result of this unavailable feature, the model cannot be used for forecasting.
In TIM, a data availability scheme is attached to each model building effort. This allows TIM to take expected data availability conditions into account when building models. Even so, it is possible that data availability changes over time or that some unexpected changes in data availability occur. Thanks to TIM’s InstantML capabilities, a model can instantly be rebuild from scratch using the constrains of this new situation. In the above example, TIM can easily build a new model without this feature, that can be used to make the desired forecast.
Situations where only a single-point forecast is required are rare in industrial practice. Multi-point forecasts are far more common, as they are required in many industrial verticals. Multi-point forecasts have traditionally been addressed by multi-output models and recurrent strategies.
Intuitively, building and optimizing a multi-output model is harder than doing so for a single-output model, because model parameters need to be optimized against all the outputs simultaneously. These models thus tend to exhibit a higher complexity compared to single-output models (e.g. more hidden layers in a neural net, a larger decision tree…). Sometimes, this can even result in a contradictory optimization problem.
Recurrent strategies, on the other hand, optimize a single-output model that is then recurrently propagated in time, i.e. the forecast for y(t+1) is reused for calculating y(t+2). Recurrent strategies are, however, prone to fast divergence. This renders them impractical for a widespread industrial adoption.
TIM addresses multi-point forecasts by creating a set of single-point models. In this way, every forecast point has its own individual model that considers the corresponding data availability constraints. This set of models is referred to as a Model Zoo. Out of this Model Zoo, TIM automatically dispatches the correct model for the calculation of each point of the multi-point forecast.
This approach has several advantages. First, by addressing each point individually, TIM can often achieve greater accuracy, because each model is optimized for a single point while considering the specific data availability. Secondly, each model’s features can be examined and compared. As there might be rather different features driving different forecasts, this can provide great insights.
Often, multiple forecasts are needed at multiple points in time. In these cases, the user defines a forecasting routine, referring to a set of forecasting situations with their corresponding data availability schemes. In these cases, TIM’s multi-situational layer assembles a Model Zoo that accounts for all the multi-point forecasts in all the situations of the user’s forecasting routine.