Class HeteroscedasticityFitDetectorOptions
Configuration options for the Heteroscedasticity Fit Detector, which analyzes whether a model's prediction errors have constant variance across all prediction values.
public class HeteroscedasticityFitDetectorOptions- Inheritance
-
HeteroscedasticityFitDetectorOptions
- Inherited Members
Remarks
Heteroscedasticity refers to the situation where the variance of errors (residuals) varies across the range of predicted values. In regression analysis, one of the key assumptions is homoscedasticity, which means the error variance should be constant across all levels of the predicted values. When this assumption is violated (heteroscedasticity), it can lead to inefficient parameter estimates and unreliable confidence intervals.
For Beginners: This detector checks if your model's errors (the differences between predicted and actual values) are consistent across all predictions or if they vary in a systematic way.
Imagine you have a model predicting house prices. If your model tends to make small errors for low-priced houses but much larger errors for expensive houses, that's heteroscedasticity. It means your model's accuracy depends on what it's predicting, which is usually not ideal.
Think of it like a weather forecaster who's very accurate when predicting mild temperatures but wildly inaccurate when predicting extreme temperatures. You'd want to know about this inconsistency because it affects how much you can trust different predictions.
When heteroscedasticity is detected, it often suggests that:
- Your model might be missing important features
- You might need to transform your target variable (e.g., use log of price instead of price)
- You might need a different type of model altogether
- You should be more cautious about the model's predictions in certain ranges
Properties
HeteroscedasticityThreshold
Gets or sets the p-value threshold for detecting heteroscedasticity in model residuals.
public double HeteroscedasticityThreshold { get; set; }Property Value
- double
The heteroscedasticity threshold, defaulting to 0.05 (5%).
Remarks
This threshold is used in statistical tests for heteroscedasticity (like the Breusch-Pagan test). If the p-value from the test is below this threshold, the null hypothesis of homoscedasticity is rejected, and the model is considered to have heteroscedastic errors. Lower values make the test more conservative, requiring stronger evidence to declare heteroscedasticity.
For Beginners: This setting determines how strong the evidence needs to be before the detector concludes that your model has inconsistent error patterns (heteroscedasticity). With the default value of 0.05, if the statistical test finds that there's less than a 5% chance that the varying error patterns happened by random chance, it will flag heteroscedasticity.
Think of it like a jury's standard of evidence - with 0.05, you're saying "I want to be at least 95% confident before declaring there's a problem with inconsistent errors." If you want to be more certain before raising an alarm, you could lower this threshold (e.g., to 0.01, requiring 99% confidence). If you want to be more sensitive to potential issues, you could raise it (e.g., to 0.1).
In statistical terms, this is a p-value threshold. The p-value represents the probability that you would see the observed pattern of errors if the errors were actually consistent (homoscedastic). A small p-value suggests that the inconsistent pattern is unlikely to be due to random chance.
HomoscedasticityThreshold
Gets or sets the p-value threshold for confirming homoscedasticity in model residuals.
public double HomoscedasticityThreshold { get; set; }Property Value
- double
The homoscedasticity threshold, defaulting to 0.1 (10%).
Remarks
This threshold is used to positively confirm homoscedasticity (constant error variance). If the p-value from the heteroscedasticity test is above this threshold, the model is considered to have homoscedastic errors. Higher values make the test more conservative, requiring stronger evidence to declare homoscedasticity.
For Beginners: This setting determines how confident the detector needs to be before declaring that your model has consistent error patterns (homoscedasticity). With the default value of 0.1, if the statistical test finds that there's more than a 10% chance that any varying error patterns happened by random chance, it will confirm homoscedasticity.
Think of it like a doctor giving you a clean bill of health - with 0.1, the doctor is saying "I'm reasonably confident you don't have this condition, but I can't be absolutely certain." This is intentionally set higher than the heteroscedasticity threshold to create a "buffer zone" between definitely heteroscedastic and definitely homoscedastic.
If the p-value falls between the heteroscedasticity threshold (0.05) and the homoscedasticity threshold (0.1), the detector will indicate that the evidence is inconclusive - there might be some inconsistency in errors, but it's not strong enough to be certain.
In practice, having consistent errors (homoscedasticity) is generally preferred because it means your model's accuracy is similar across all predictions, making it more reliable and easier to interpret.