Time-penalised tree: an interpretable temporal algorithm applied to climate risks[Time-penalised tree : un algorithme temporel interprétable appliqué aux risques climatiques]

This presentation introduces an innovative decision tree algorithm designed to take into account temporal variables in dynamic environments. Traditional methods often fail when the explanatory variables vary over time, leading to inaccurate forecasts. Our solution, the Time-penalised Tree (TpT) algorithm, uses a time-penalised division criterion, allowing the joint recursive partitioning of the covariate space and time. This approach incorporates historical trends into the construction of the model, providing a clear and interpretable framework. We present the structure and operation of the TpT algorithm, highlighting its advantages over existing methods. A case study on climate risk prediction illustrates the practical application of TpT, showing how it improves the interpretability of forecasts using historical climate data. Since these data form time series, in space and in time, the TpT algorithm also helps in the selection of the most predictive climate variables. We also discuss the theoretical properties of TpT, its effectiveness and its potential application in various fields such as health, finance and insurance. Future research directions are discussed, including the validation and comparison of TpT with other algorithms on various actuarial datasets.