Working version of NeuralProphet with missing features and potentially containing bugs.
Modelling capabilities:
- [done] Trend, piecewise linear changepoints
- [done] Auto-regression, univariate, multi-step ahead forecasts
- [done] Seasonalities, based on fourier-terms
- [done] Optional hidden layers for AR
- [done] Manage missing data - basic automatic imputation
- [done] Basic Automatic hyperparameter selection
- [done] Custom Metrics
- [done] Training with evaluation on holdout set
- [done] Events and Holidays
- [done] Exagenous variables (as covariate inputs)
- Simple Uncertainty estimation
User Interface:
- simple package with limited capabilities
- similar to Facebook Prophet's basic features
Accompanying Products:
- Quickstart documentation and examples
- Benchmarks (Accuracy and Execution time)
- Few datasets
Added modelling capabilities:
- More intelligent Automatic hyperparameter selection
- different ways to manage trend/normalize data and compute seasonality (rolling, local seasonality, ...)
- Inclusion of traditional models (ets, sarimax, ...)
- Component-wise uncertainty
User Interface:
- More user-control (set trend changepoint times, ...)
- Rich analytics and plotting
- Model gives user feedback on how to improve hyperparameters (if set)
- Integration with Time-Series Preprocessing tools
Accompanying Products:
- Large collection of time-series datasets
- Professional documentation and more tutorials
Here, we will re-write large portions of the code structure in order to make it a modular framework where model components can freely be interchanged and combined.
Added modelling capabilities:
- Inclusion of more potent models (Recurrence, Convolution, Attention, ...)
User Interface:
- Tools for Understanding of model and input-output mapping
- Integration with relevant Interfaces (Pytorch metrics, Tensorboard, scikitlearn, ...)
Accompanying Products:
- Pre-trained models
Alternative visual web-interface, potentially cloud-based execution