Collect Ideas for the PyAF 5.0 roadmap

[antoinecarme]

This is an issue to serve as a honeypot for all ideas that come and can be included in the next PyAF release (5.0, expected on 2023-07-14).

First ideas :

1. Forecasting Models based on Deep Learning Attention Mechanisms
2. Recurrent tasks : More hardware architectures, forecasting performance measures, benchmarks, bug fixes, docs and optimization/profiling.
3. User feedback and issues

Ideally, the main part of the implementation will take place during the coming covid lockdown (Fall 2022 ;-).

[antoinecarme]

Large Horizon Models (H large enough). Profiling for CPU/memory/speed.

#213

[antoinecarme]

Model Esthetics

#212

[antoinecarme]

Get rid of Keras and tensorflow. Use PyTorch. Less is better. Minimal is beautiful.

#211

[antoinecarme]

PyAF will use PyTorch as its deep learning architecture for future projects.

#211

[antoinecarme]

Investigate business days / hours impact.

#210

[antoinecarme]

Add some outlier-resistant forecasting performance measures for robustness.

#209

[antoinecarme]

New Hardware Architecture : RISC-V

#208

[antoinecarme]

Investigate Threshold AR (TAR) Models

#214

[antoinecarme]

Investigate TSMARS Models

#215

[antoinecarme]

#221

[antoinecarme]

Implemented/Processed So far :

2. RISC-V Hardware Platform Validation #208
3. Outlier-resistant forecasting Performance Measures #209
4. Use PyTorch as the reference deep learning framework/architecture for future projects #211
5. Investigate Model Esthetics for PyAF #212
6. Investigate Large Horizon Models #213
7. Bad plot for shaded area around prediction intervals in hourly data #216
8. Automate Prototyping Activities - R-based Models #217
9. Failure to build a multiplicative ozone model with Lag1 trend #220
10. Add Differentiable Variant of SMAPE Performance Measure #221
11. Revisit Model Complexity Definition #223
12. Run some Sanity Checks for PyAF 5.0 #224
13. Forecast Quantiles Plots can be improved  #225

[antoinecarme]

Release Process starts.

#228

CLOSING