Aidenzich
commented
3 months ago - https://arxiv.org/pdf/2403.07815.pdf
| Why (Problem/Need) | What (Solution/Tool) | How (Method/Approach) |
|---|---|---|
| Traditional time series forecasting has been limited by models being trained and predicted on the same dataset, with a recent shift towards deep learning methods due to the availability of diverse data sources. However, these methods still face limitations in terms of generalization across different datasets and computational efficiency. | Chronos is introduced as a simple yet effective framework for pretraining probabilistic time series models that can leverage the strengths of transformer-based language model architectures without significant modifications. | Chronos tokenizes time series data using scaling and quantization, transforming real-valued time series into a discrete token sequence. These tokens are then used to train transformer-based models using a cross-entropy loss function, allowing the model to learn the sequential structure of time series data for forecasting. |
| There's a scarcity of publicly available time series datasets in terms of quantity and quality, which is crucial for developing general-purpose forecasting models that can perform well in zero-shot scenarios across diverse domains. | To enhance the diversity and quality of training data for Chronos, data augmentation techniques like TSMix (Time Series Mixup) and KernelSynth (Synthetic Data Generation using Gaussian Processes) are integrated into the training process. | TSMix creates new time series by taking convex combinations of randomly sampled base time series from the training datasets, while KernelSynth generates synthetic time series by composing kernel functions with Gaussian processes. These strategies help overcome the limitations of small training datasets, improving model robustness and generalization. |