Inspired by recent advances in large language models, foundation models have been developed for zero-shot time series forecasting, enabling prediction on datasets unseen during pretraining. These large-scale models, trained on vast collections of time series, learn generalizable representations for both point and probabilistic forecasting, reducing the need for task-specific architectures and manual tuning. In this work, we review the main architectures, pretraining strategies, and optimization methods used in such models, and study the effect of fine-tuning after pretraining to enhance their performance on specific datasets. Our empirical results show that fine-tuning generally improves zero-shot forecasting capabilities, especially for long-term horizons.

Real-time energy forecasting on edge devices represents a major challenge for smart grid optimization and intelligent buildings. We present LAD-BNet (Lag-Aware Dual-Branch Network), an innovative neural architecture optimized for edge inference with Google Coral TPU. Our hybrid approach combines a branch dedicated to explicit exploitation of temporal lags with a Temporal Convolutional Network (TCN) featuring dilated convolutions, enabling simultaneous capture of short and long-term dependencies. Tested on real energy consumption data with 10-minute temporal resolution, LAD-BNet achieves 14.49% MAPE at 1-hour horizon with only 18ms inference time on Edge TPU, representing an 8-12 x acceleration compared to CPU. The multi-scale architecture enables predictions up to 12 hours with controlled performance degradation. Our model demonstrates a 2.39% improvement over LSTM baselines and 3.04% over pure TCN architectures, while maintaining a 180MB memory footprint suitable for embedded device constraints. These results pave the way for industrial applications in real-time energy optimization, demand management, and operational planning.

The Recursive KalmanNet, recently introduced by the authors, is a recurrent neural network guided by a Kalman filter, capable of estimating the state variables and error covariance of stochastic dynamic systems from noisy measurements, without prior knowledge of the noise characteristics. This paper explores its generalization capabilities in out-of-distribution scenarios, where the temporal dynamics of the test measurements differ from those encountered during training. Le Recursive KalmanNet, récemment introduit par les auteurs, est un réseau de neurones récurrent guidé par un filtre de Kalman, capable d'estimer les variables d'état et la covariance des erreurs des systèmes dynamiques stochastiques à partir de mesures bruitées, sans connaissance préalable des caractéristiques des bruits. Cet article explore ses capacités de généralisation dans des scénarios hors distribution, où les dynamiques temporelles des mesures de test diffèrent de celles rencontrées à l'entraînement.

In healthcare, patient data is often collected as multivariate time series, providing a comprehensive view of a patient's health status over time. While this data can be sparse, connected devices may enhance its frequency. The goal is to create patient profiles from these time series. In the absence of labels, a predictive model can be used to predict future values while forming a latent cluster space, evaluated based on predictive performance. We compare two models on Withing's datasets, M AGMAC LUST which clusters entire time series and DGM${}^2$ which allows the group affiliation of an individual to change over time (dynamic clustering).

This paper presents a new algorithm: TSRuleGrowth, looking for partially-ordered rules over a time series. This algorithm takes principles from the state of the art of rule mining and applies them to time series via a new notion of support. We apply this algorithm to real data from a connected environment, which extract user habits through different connected objects.

The analysis of the behaviour of individuals and entities (UEBA) is an area of artificial intelligence that detects hostile actions (e.g. attacks, fraud, influence, poisoning) due to the unusual nature of observed events, by affixing to a signature-based operation. A UEBA process usually involves two phases, learning and inference. Intrusion detection systems (IDS) available still suffer from bias, including over-simplification of problems, underexploitation of the AI potential, insufficient consideration of the temporality of events, and perfectible management of the memory cycle of behaviours. In addition, while an alert generated by a signature-based IDS can refer to the signature on which the detection is based, the IDS in the UEBA domain produce results, often associated with a score, whose explainable character is less obvious. Our unsupervised approach is to enrich this process by adding a third phase to correlate events (incongruities, weak signals) that are presumed to be linked together, with the benefit of a reduction of false positives and negatives. We also seek to avoid a so-called "boiled frog" bias inherent in continuous learning. Our first results are interesting and have an explainable character, both on synthetic and real data.

This paper proposes a method of segmenting temporal data into ordered classes. It is based on mixture models and a discrete latent process, which enables to successively activates the classes. The classification can be performed by maximizing the likelihood via the EM algorithm or by simultaneously optimizing the model parameters and the partition by the CEM algorithm. These two algorithms can be seen as alternatives to Fisher's algorithm, which improve its computing time.