Time-Series Prediction Platform
Time-series prediction is a common problem in multiple domains for a variety of applications, including retail, industry, smart cities, and financial services. The span of research in the time series field is growing exponentially, with hundreds of deep learning time-series forecasting paper submissions to ICML, ECML, ITISE, and multiple journals every year. However, there is currently no common benchmark framework to compare the accuracy and performance of all the models from industry or academia.
The Time-Series Prediction Platform enables users to easily mix and match datasets and models. In that case, the user has full control over the following settings, being able to compare side-by-side results obtained from various solutions. Those, include:
- Evaluation metrics
- Evaluation datasets
- Prediction horizons
- Prediction sliding window sizes
The Time-Series Prediction Platform has a fully modular and configurable architecture and inbuilt Automatic Mixed Precision and Multi-GPU support, which enables rapid iteration and experimentation. The platform also provides functionality to automatically schedule multiple runs, capture results in an easily digestible format, and run hyperparameter search with Optuna.
Examples used are the standard benchmark datasets, baseline model, ARIMA model, Temporal Fusion Transformer, and more to come.