Selffitting

Selffitting, also written self-fitting, is the process by which a statistical or computational model automatically selects or estimates its parameters and, in some cases, its model form, with limited or no manual tuning. The goal is to adapt the model to the data in a data-driven way, reducing expert intervention and increasing reproducibility.

Common approaches combine parameter estimation with model selection. Estimation often uses maximum likelihood or least squares, while model selection relies on cross-validation, information criteria (such as AIC or BIC), regularization, or Bayesian methods. Regularization introduces penalties to curb overfitting and can aid in automatic determination of model complexity. Bayesian techniques can infer parameter values and, in hierarchical or nonparametric variants, adapt aspects of the model structure itself.

Typical workflows include preprocessing the data, defining a fitting objective, performing automatic estimation, evaluating fit and predictive performance, and selecting a final model. In practice, selffitting is used in statistics, signal processing, time-series analysis, and machine learning, and underpins ideas behind automatic model selection and autoML.

Advantages of selffitting include improved reproducibility, reduced manual trial-and-error, and the ability to handle complex or high-dimensional data. Limitations include computational cost, the risk of overfitting or selecting overly complex models if not properly regularized, and potential opacity of the resulting model choices.

Terminology varies: some communities prefer “self-fitting” or refer to related concepts as automatic model selection, automatic tuning, or autoML. The core idea is data-driven adaptation of models without extensive manual parameter tuning.