Model Masks

Types of Masks and Intent

Watchtower uses 3 types of masks for modeling purpose: 1. no_prediction_mask: To identify the points which should not be trained nor predicted 2. no_train_mask: To identify the points which should note be trained, but due to being causal outliers should be predicted. The mask will not have its prediction errors attributed to any driver 3. feature_masks: To identify the points which should note be trained, but due to known externalities should be predicted. Each mask will attribute the prediction error to a different driver

NOTE: Points in no_train_mask or feature_masks which overlap with no_prediction_mask will not have predictions

What is included?

1. no_prediction_mask: Points of non-positive sales or non-positive price or miniscule distribution
2. no_train_mask: Points of sales below a particular threshold pre 2021 or last few weeks of sales
3. feature_masks: Points where the sales spikes or plummets. They can be of two types:
   • Festivals: To identify points where the sales spiked because of certain events or holidays
   • Supply Chain Mask: To idenfiy points where the sales plummeted due to issues in the supply chain

Driver Attribution in Feature Masks

For masked datapoints: 1. Make Smoothed Sales same as True Sales 2. Generate model predictions 3. Calculate the difference between predictions and True Sales. Attribute this difference to Festivals or Supply Chain Issues 4. Make the predictions from step 2 same as True Sales and call it Predicted Sales