[Major] Glocal Modelling v2

[alfonsogarciadecorral]

WORK IN PROGRESS

This PR will include a new set of features for Global-Local Modelling:

• Glocal Trend - To capture both global and local trend information.

A test solution is already implemented. We have to translate. Will be done after Trend Modularity PR is pushed

• Different Seasonality modes - To allow using a global yearly seasonality and a local weekly seasonality for example
• Glocal Seasonality - To capture both global and local seasonality information.
• Static Covariates - To allow local modelling of regressors/covariates/AR.
• Future Regressors - Better modelling of the regressors impact

Once the implementation is done. We will test the model using retail data.

[github-actions]

7ee51ed

Model Benchmark

Benchmark	Metric	main	current	diff
AirPassengers	MAE_val	15.2698	15.2698	0.0%	✅
AirPassengers	RMSE_val	19.4209	19.4209	0.0%	✅
AirPassengers	Loss_val	0.00195	0.00195	0.0%	✅
AirPassengers	MAE	9.82902	9.82902	0.0%	✅
AirPassengers	RMSE	11.7005	11.7005	0.0%	✅
AirPassengers	Loss	0.00056	0.00056	0.0%	✅
AirPassengers	time	4.2567	4.4	3.37%	⚠️
PeytonManning	MAE_val	0.64636	0.64636	0.0%	✅
PeytonManning	RMSE_val	0.79276	0.79276	0.0%	✅
PeytonManning	Loss_val	0.01494	0.01494	0.0%	✅
PeytonManning	MAE	0.42701	0.42701	0.0%	✅
PeytonManning	RMSE	0.57032	0.57032	0.0%	✅
PeytonManning	Loss	0.00635	0.00635	0.0%	✅
PeytonManning	time	11.9637	12.07	0.89%	✅
YosemiteTemps	MAE_val	1.72949	1.72949	0.0%	✅
YosemiteTemps	RMSE_val	2.27386	2.27386	0.0%	✅
YosemiteTemps	Loss_val	0.00096	0.00096	0.0%	✅
YosemiteTemps	MAE	1.45189	1.45189	0.0%	✅
YosemiteTemps	RMSE	2.16631	2.16631	0.0%	✅
YosemiteTemps	Loss	0.00066	0.00066	0.0%	✅
YosemiteTemps	time	92.301	100.23	8.59%	❌

Model training plots

Model Training

PeytonManning

YosemiteTemps

AirPassengers

[github-actions]

Model Benchmark

Benchmark	Metric	main	current	diff
YosemiteTemps	MAE_val	0.56412	0.56412	0.0%	✅
YosemiteTemps	RMSE_val	0.83161	0.83161	0.0%	✅
YosemiteTemps	Loss_val	0.0004	0.0004	0.0%	✅
YosemiteTemps	MAE	0.98449	0.98449	0.0%	✅
YosemiteTemps	RMSE	1.75389	1.75389	0.0%	✅
YosemiteTemps	Loss	0.00132	0.00132	0.0%	✅
YosemiteTemps	time	34.9044	37.45	7.29%	❌
EnergyPriceDaily	MAE_val	5.41157	5.41157	0.0%	✅
EnergyPriceDaily	RMSE_val	6.71538	6.71538	0.0%	✅
EnergyPriceDaily	Loss_val	0.02525	0.02525	0.0%	✅
EnergyPriceDaily	MAE	5.94936	5.94936	0.0%	✅
EnergyPriceDaily	RMSE	7.9833	7.9833	0.0%	✅
EnergyPriceDaily	Loss	0.02579	0.02579	0.0%	✅
EnergyPriceDaily	time	15.3339	16.06	4.74%	⚠️
PeytonManning	MAE_val	0.34955	0.34955	0.0%	✅
PeytonManning	RMSE_val	0.50049	0.50049	0.0%	✅
PeytonManning	Loss_val	0.01771	0.01771	0.0%	✅
PeytonManning	MAE	0.34653	0.34653	0.0%	✅
PeytonManning	RMSE	0.49312	0.49312	0.0%	✅
PeytonManning	Loss	0.01464	0.01464	0.0%	✅
PeytonManning	time	12.6275	12.78	1.21%	✅
AirPassengers	MAE_val	30.6285	30.6285	0.0%	✅
AirPassengers	RMSE_val	31.5254	31.5254	0.0%	✅
AirPassengers	Loss_val	0.01277	0.01277	0.0%	✅
AirPassengers	MAE	6.16861	6.16861	0.0%	✅
AirPassengers	RMSE	7.85225	7.85225	0.0%	✅
AirPassengers	Loss	0.00065	0.00065	0.0%	✅
AirPassengers	time	8.21262	8.81	7.27%	❌

Model training plots

Model Training

PeytonManning

YosemiteTemps

AirPassengers

EnergyPriceDaily

[ourownstory]

Excellent Work @alfonsogarciadecorral !!
This is a massive contribution and a big upgrade for global-local modeling.
Thank you!

[ourownstory]

@leoniewgnr Can you please help @alfonsogarciadecorral to get the tests to run, so we can merge this large PR? The issue does not seem connected to this PR, but rather to poetry? @hxyue1 Can you maybe help figure out if this is a poetry misconfiguration? Thank you so much!

[codecov]

Codecov Report

Attention: 13 lines in your changes are missing coverage. Please review.

Comparison is base (a99059a) 88.40% compared to head (c4b04e4) 88.51%.

Files	Patch %	Lines
...components/future_regressors/shared_neural_nets.py	90.24%	4 Missing ⚠️
...nents/future_regressors/shared_neural_nets_coef.py	90.47%	4 Missing ⚠️
...rophet/components/future_regressors/neural_nets.py	95.83%	2 Missing ⚠️
neuralprophet/configure.py	94.44%	2 Missing ⚠️
neuralprophet/components/router.py	95.83%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #1008      +/-   ##
==========================================
+ Coverage   88.40%   88.51%   +0.11%     
==========================================
  Files          38       41       +3     
  Lines        5105     5329     +224     
==========================================
+ Hits         4513     4717     +204     
- Misses        592      612      +20     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[alfonsogarciadecorral]

Hi @leoniewgnr @hxyue1 ! Just saw this is still not merged, let me know if you have any doubt about anything from the PR! :)

[ourownstory]

@alfonsogarciadecorral I finally got to do a deep-dive and found the bug: A model forward pass in time_net was computing AR components twice - once directly and once after stationarizing. Now the performance metrics match!

[ourownstory]

• Next steps: create new release before merging
• then merge and create another release with this.