# Forecasting Roadmap 2025-26

Forecasting is focusing on expanding the range of algorithms in addition to expanding the use cases.

1. We will exclusively use numpy arrays (rather than dataframes) until we find a need to adapt.
2. We will keep the interface as simple as possible, concentrate on benchmarked bespoke implementations that are based on numba.
3. We will try to engage authors of algorithms we are implementing

## Structure

1. Finalise iterative/direct interface, possible third one for ETS type #2899
2. Make y compulsory for predict, make naive and ets fit is empty #2927
3. Implement as mixins rather than base class methods
4. Facilitate multivariate forecasting with first example
5. Restructure into stats/ml/deep learning
6. Allow series to series forecasters in the framework #2910

## Stats forecasters

1. ARIMA #2860
2. AR only
3. SARIMA #2862
4. AutoARIMA #2861
5. AutoETS #2828, #2411
6. TBATS #2847
7. Threhold AR models #2817
8. STL #2851 and MSTL #2849
9. Theta Forecaster

## Machine learning forecasters

1. Allow exogenous in RegressionForecaster #2915
2. implement SETAR-Forest #2816
3. DONE: TVP #2920

## Deep forecasters
1. N-BEATS #2850
2.

## Benchmarking
1. Compare stats forecasters on simulated ARMA data
2.