A **model fitting norm** is an approach for evaluating the "goodness of fit" of predictive models (e.g. [[Linear Regression]]) to the underlying data. There is no single, best model-fitting norm, but some common examples are:
- [[Mean Squared Error]] (MSE)
- Root MSE
- Lasso
- Ridge
- Elastic Net
To apply these norms, you take each of the observed and predicted values and compute their error. Then sum the errors of all terms. Because errors between `observed` and `predicted` values can be positive or negative, and [[Errors add, not cancel]], we take steps to mitigate cancellation.
Either:
- **Absolute error** - taking the absolute value of the error
- **Square error** - taking the square of the error
- L1 (**Manhattan**) norm uses absolute error
- This is called "manhattan" because it's like walking on vertical and horizontal tracks to find your error
- L2 (**Euclidean**) norm uses square error
- This takes the _hypotenuse_ to find the error
- L`k` norm uses the `k`th exponent
****
# More
## Source
- Grad School
## Related
- [[R2]]