07. Model Drift and Drift Detection Techniques
AI For Trading C6 L4 A06 Model Drift And -Drift Detection Techniques V2
Understanding Model Performance in Financial Domains
Essentials of Model Evaluation:
- Historical Data Utilization:
- Used for training and evaluating predictive models.
- "Past performance is not indicative of future results." Reminder of dynamic markets.
- Alpha Decay:
- Occurs when models lose predictive ability over time.
Monitoring Model Efficiency:
- Model Drift Awareness:
- Data Drift: Changes in data distribution
- Concept Drift: Changes in the relationships between data features and targets
- Continuous Monitoring:
- Evaluate ongoing performance to detect drift and preempt potential losses
Statistical Tools and Tests for Detection:
- Kolmogorov-Smirnov (K-S) Test:
- Analyses continuous-valued feature shifts.
- Sensitive to central distribution changes.
- Chi-squared Test:
- Suitable for categorical features.
- Better detects changes in distribution tails.
Regular implementation of these tests ensures early detection of model degradation, thereby safeguarding performance longevity.
QUIZ QUESTION::
Match the concepts related to data science and machine learning with their definitions.
ANSWER CHOICES:
|
Definitions |
Concepts |
|---|---|
The phenomenon where the accuracy of a machine learning model decreases over time due to changes in the underlying data distribution. |
|
The change in the relationship between input data and the target variable that can lead to decreased performance. |
|
The shift in the input data distribution that occurs over time, while the underlying relationship with the target remains the same. |
SOLUTION:
|
Definitions |
Concepts |
|---|---|
|
The shift in the input data distribution that occurs over time, while the underlying relationship with the target remains the same. |
|
|
The change in the relationship between input data and the target variable that can lead to decreased performance. |
|
|
The phenomenon where the accuracy of a machine learning model decreases over time due to changes in the underlying data distribution. |
SOLUTION:
- Compare the distributions of numerical features at various points in time to detect shifts
- Compare the distributions of model predictions at two different time points to detect shifts