Nixtla has announced the launch of NeuralForecast, an advanced library of neural forecasting models set to revolutionise the forecasting community. The library addresses long-standing issues such as usability, accuracy, and computational efficiency, providing a bridge between neural networks’ complexity and their practical use.
NeuralForecast comprises multiple neural network architectures, from Multi-Layer Perceptrons (MLP) and Recurrent Neural Networks (RNNs) to more complex models like NBEATS, NHITS, Temporal Convolutional Networks (TCNs), Temporal Fusion Transformer (TFT), and Informer. This suite of models caters to a wide range of forecasting needs, offering users access to state-of-the-art techniques.
Key features of NeuralForecast include usability and robustness which is achieved by offering a unified interface compatible with popular forecasting libraries like StatsForecast and MLForecast. This compatibility streamlines workflows, enabling seamless transitions between libraries and enhancing productivity. It also supports static, historical, and future exogenous variables, providing flexibility in model inputs and enabling the incorporation of external factors into forecasting models to improve accuracy.
Furthermore, NeuralForecast facilitates forecast interpretability by offering tools to interpret forecasts by plotting trend, seasonality, and exogenous prediction components. This capability aids users in understanding the underpinning patterns and influences in their data. It also delivers probabilistic forecasting by offering simple model adapters for quantile losses and parametric distributions. This functionality allows users to generate forecasts with confidence intervals, providing a more comprehensive understanding of potential future outcomes.
Efficient automatic model selection is another key feature of NeuralForecast. The library includes parallelized automatic hyperparameter tuning, efficiently searching for the best validation configuration. This feature can significantly cut down the time and computational resources needed for model optimization.
NeuralForecast’s launch marks a significant development in the application of neural networks in forecasting, addressing core challenges in usability, robustness, and leveraging state-of-the-art models. Hence, the library could become an indispensable tool for data scientists and forecasters seeking to fully utilize neural networks.
As an illustration of its usage, the announcement includes a sample code showing how to use NeuralForecast with NBEATS and NHITS models to forecast monthly passenger data.
In a nutshell, Nixtla’s NeuralForecast library promises to create a significant impact in the field of forecasting, addressing critical barriers and providing a user-friendly platform to optimize neural networks’ application in prediction models.
