Why Data Scientists Get Blamed for Model Failures (Even When the Model Isn’t the Problem)

Read time: 2.5 minutes

The Real Reason Data Scientists Take the Blame

1️⃣ The Success of Working Models is Not Being Acknowledged
Models run into production when they are successful; however, their contributions disappear once they become part of everyday operations.
Solution: Share a Model Impact Scorecard on a monthly basis, providing revenue generated, losses avoided and who/how many decisions were altered.

2️⃣ Technical Metrics Have Little to No Business Value
Executive teams do not care whether a model has AUC, F1, or Precision... it does not provide them any value.
Solution: Provide a dollar amount for each metric to link back to either reduced risk or operational impact.

3️⃣ Many so-called Model Failures Are Actually Data Failures
Broken pipelines, incorrect labels, and data drift are the primary reasons for model failure.
Solution: Monitor and track the features that drift, labels that drift and predictions that drift to quickly identify the root cause of failure.

4️⃣ Black Box Models Cause Distrust
Trust is diminished when decision makers lack understanding of the model under scrutiny.
Solution: Add capabilities to provide Explainability to models by using SHAP values, Feature Importance, and Driver Analysis.

5️⃣ Lost Budget from a Lack of Visibility Into Model Performance
Because stakeholders cannot see the value of a model, funding will no longer be available.
Solution: Create a Model Value Dashboard to show the number of predictions, the number of decisions that have been performed automatically, and the financial impact of doing so.

💡Key Takeaway: 

Models will not just go away with a whimper. The success of models that are not obvious will ultimately become a reason to blame.

If you want your model to be successful in the real world, it should be impossible to ignore.