Desired Standard Deviations

Description

Desired Standard Deviation defines the strength of the regularization loss in the hierarchical embedding layer. Higher the desired standard deviation, lower is the hierarchical regularization loss allowing the hierarchical embeddings to take more deviated values.

Metrics

1) Combined R2 based on combined validation periods of WFO

Validation Period - 2022 Q2, Q3, Q4 with 512 epochs and batch size of 512

Desired Standard Deviations and Combined R2 Metric

Global ME

Global ME Combined R2

It is observed that lowering the parameter below current value of 10.0 is beneficial.

Holiday ME

Holiday ME Combined R2

It is observed that increasing the parameter above current value of 0.1 is beneficial.

Distribution

Distribution Combined R2

It is observed that the current value is optimal.

Pricing ME

Pricing ME Combined R2

It is observed that the current value is optimal.

Invest ROI - Sigmoid and Betagamma

Invest ROI Combined R2

No discernible pattern observed in 3D plot. No action taken.

Pricing

Pricing Combined R2

Parameter value left unchanged.

Promo ROI Sigmoid Combined R2

Sigmoid related parameter left unchanged.

Promo ROI Betgamma Combined R2

It is observed that increasing the betagamma related parameter above the current value of 0.1 is beneficial.

Comparison of Metrics based Original and Updated Hyperparamters

Following link provides wfo metrics of models with original and updated hyperparameters.

Metrics - Original vs Updated Hyperparameters*

Improvement in combined R2 metric is observed with the updated hyperparameters.

*Generated on 27th June 2023