Time series forecasting has become an essential element for decision-making in various industries, from predicting traffic to estimating sales. These prediction capabilities allow organizations to make informed decisions, mitigate risks, and allocate resources more efficiently. However, traditional machine learning methods often require extensive adjustments and customization, prolonging development and consuming a large amount of resources.
In this context, innovations such as Chronos emerge, a family of time series models that leverage the power of large language model (LLM) architectures to overcome these limitations. Chronos has been pre-trained with large and diverse datasets, allowing it to generalize its forecasting ability across multiple domains. This feature allows it to excel in “zero-shot” forecasts, making predictions without being specifically trained on the target dataset, outperforming most task-specific models in evaluations.
Chronos is based on the observation that both LLMs and time series forecasting seek to decode sequential patterns to predict future events. This similarity allows treating time series data as a language to model using transformer architectures. The model converts continuous time series data into a discrete vocabulary through a two-stage process: scaling the data and quantizing it into a fixed number of equidistant bins.
An integration of Chronos with Amazon SageMaker Pipeline is foreseen, using a synthetic dataset that simulates a sales forecasting scenario. This integration will unlock accurate and efficient predictions with minimal data. Users will be able to learn to use features that coordinate the entire workflow, from tuning to deployment, thus optimizing the development process and enabling the application of Chronos to any time series dataset.
For those interested in this process, access to a SageMaker domain with the required AWS Identity and Access Management permissions for resource creation and management is necessary. SageMaker Pipelines will facilitate the orchestration of training and evaluation experiments, allowing for the execution of multiple iterations simultaneously and reducing both processing time and costs.
Once the forecasting model is completed, it will be deployed using SageMaker’s hosting services, creating an endpoint for making real-time predictions. This will allow seamless integration with existing applications and systems, offering on-demand access to the predictive capabilities of the model through a secure HTTPS interface.
Chronos’ performance has been evaluated on 27 datasets not used during training, demonstrating its effectiveness in zero-shot predictions compared to local statistical and task-specific models. The results highlight Chronos’ ability to generalize and adapt to previously unaddressed situations.
This innovation represents a significant advancement for companies in various sectors, by facilitating the implementation of advanced time series forecasting techniques without the need for extensive in-house machine learning expertise, significantly enhancing decision-making and operational efficiency.
Source: MiMub in Spanish
