paper.md

title	LazyPredict: Rapid Multi-Model Benchmarking for Machine Learning and Time Series Forecasting in Python
tags	Python machine learning automated machine learning model selection time series forecasting scikit-learn
authors	name orcid affiliation Shankar Rao Pandala 0009-0006-2234-6944 1
affiliations	name index Independent Researcher 1
date	15 April 2026
bibliography	paper.bib

Summary

LazyPredict is a Python library that enables rapid benchmarking of machine learning models with minimal code. Given a dataset, it automatically fits and evaluates over 40 classification and regression models from scikit-learn and popular gradient boosting libraries, returning a ranked comparison table of performance metrics. The library extends this approach to time series forecasting with 26+ models spanning statistical, machine learning, deep learning, and foundation model categories. LazyPredict provides researchers and practitioners with a systematic, reproducible way to identify promising model families before investing effort in hyperparameter tuning, enabling faster iteration in the model selection phase of machine learning workflows.

Statement of Need

Model selection is a critical early step in any machine learning project. Practitioners typically face a large landscape of candidate algorithms and must decide which model families warrant further investigation. This process is often ad hoc: researchers manually instantiate a handful of models, fit them one by one, and compare metrics in a spreadsheet or notebook. This approach is tedious, error-prone, and frequently biased toward models the practitioner already knows.

Existing AutoML frameworks such as Auto-sklearn [@feurer2015], TPOT [@olson2016], PyCaret [@pycaret2020], H2O AutoML [@h2o2024], and FLAML [@wu2021] address a broader optimization problem: they search over model types and hyperparameters simultaneously, often requiring substantial compute time. While powerful, these tools may be more than what is needed when the goal is simply to survey which model families perform well on a given dataset before committing to a full tuning pipeline.

LazyPredict fills this gap by focusing exclusively on rapid, zero-configuration model benchmarking. With two lines of code, a user obtains a comprehensive comparison of all applicable models using their default hyperparameters. This lightweight approach serves several use cases in scientific computing:

• Exploratory analysis: Quickly identifying which model families suit a new dataset before investing in hyperparameter optimization.
• Baseline establishment: Generating a reproducible set of baseline results against which tuned or custom models can be compared.
• Pedagogical use: Teaching students about the diversity of machine learning algorithms and their relative strengths on different data characteristics.
• Time series model surveys: Benchmarking 26+ forecasting models, including statistical methods (ETS, ARIMA, Theta), ML regressors with lag features, deep learning (LSTM, GRU), and foundation models (TimesFM), in a single call.

LazyPredict has been adopted by the research community, accumulating over 45 citations in Google Scholar across diverse domains. Examples include heart failure prediction in biomedical informatics [@mahgoub2023], spam detection using word embeddings [@fellah2024], sepsis disease prediction [@alhashimi2023], and thermal conductivity modeling in chemical engineering [@varzandeh2025]. The library is available on both PyPI and conda-forge and supports Python 3.9 through 3.14.

Software Design and Features

LazyPredict is built around a modular architecture with three main estimator classes that share a common base class (LazyEstimator):

• LazyClassifier: Benchmarks classification models, reporting accuracy, balanced accuracy, ROC AUC, and F1 score for each.
• LazyRegressor: Benchmarks regression models, reporting adjusted R-squared, R-squared, and RMSE.
• LazyForecaster: Benchmarks time series forecasting models, reporting MAE, RMSE, MAPE, SMAPE, MASE, and R-squared.

The library handles preprocessing automatically, including detection and encoding of categorical features via configurable strategies (one-hot, ordinal, target, and binary encoding). A scikit-learn Pipeline wraps each model with the appropriate preprocessor, ensuring that all transformations are applied consistently.

Key architectural decisions include:

• Graceful degradation: Models that fail to fit (due to convergence issues, incompatible data, or timeouts) are caught and skipped without interrupting the full benchmark. Per-model timeout enforcement prevents any single model from blocking the run.
• Optional dependency management: Heavy dependencies (XGBoost, LightGBM, CatBoost, PyTorch, statsmodels, pmdarima, SHAP, MLflow) are optional extras, keeping the core installation lightweight.
• GPU acceleration: Transparent GPU support for XGBoost, LightGBM, CatBoost, cuML (RAPIDS), LSTM/GRU (PyTorch CUDA), and TimesFM, with automatic CPU fallback.
• Extensibility: Users can pass custom model lists, custom metrics, and configure cross-validation folds, enabling adaptation to domain-specific evaluation protocols.

Additional modules provide hyperparameter tuning via Optuna and scikit-learn search (tuning), model explainability through permutation importance and SHAP (explainability), ensemble construction from top-performing models (ensemble), and MLflow experiment tracking (integrations.mlflow).

Comparison with Related Software

Feature	LazyPredict	Auto-sklearn	TPOT	PyCaret	FLAML
Zero-config benchmarking	Yes	No	No	Partial	No
Lines of code to compare 40+ models	2	10+	10+	5+	10+
Time series forecasting	Yes (26+ models)	No	No	Yes	Partial
Hyperparameter tuning	Optional	Core	Core	Core	Core
GPU acceleration	Yes	No	No	Partial	No
Default install size	Minimal	Heavy	Heavy	Heavy	Moderate

LazyPredict is intentionally complementary to full AutoML pipelines: it is designed for the survey phase of model selection, after which users may employ more compute-intensive tools for the optimization phase.

Acknowledgements

We thank the contributors to the LazyPredict project, including Breno Batista da Silva, Murat Toprak, Chanin Nantasenamat, and all community members who have submitted issues and pull requests. We also acknowledge the developers of scikit-learn, pandas, NumPy, and the broader Python scientific computing ecosystem on which LazyPredict depends.

AI Disclosure

Generative AI tools (GitHub Copilot, Claude) were used during development to assist with code generation, documentation drafting, and test writing. All AI-generated content was reviewed, tested, and validated by the author.

References