Autonomous Feature Engineering Agent for Time-Series Tabular Data

Abstract

Existing time-series datasets from most industries contain complex sequential patterns that are difficult to analyze with conventional machine learning methods. These methods are heavily reliant on manual feature engineering. This approach is time-consuming, domain-critical, resource-intensive, and is at risk of temporal data leakage. This study introduces AutoFeat Agent, a system for automated feature engineering and classification for time-series forecasting using the tsfresh library. This system provides a complete feature extraction pipeline, including the ingestion of data, preprocessing, rolling-window feature extraction, statistical model training, evaluation, and the visualization of results. The integrated framework combines automated statistical feature extraction and engineered features to augment the prediction capability, while still maintaining temporal aspects of the data. A rolling-window approach prevents leakage in the system, which traditional feature engineering is prone to. Features are screened using the False Discovery Rate (FDR) to retain features of significance. Br os, 2nd ed. The system gives a variety of machine learning models to choose from, such as Grad Boost, RF, and Log Classification. Testing shows the system successfully reduced approximately 4698 extracted features to nearly 460, which still retain their classification capabilities and accuracy (75% to 88%) evaluating predictive models formed with financial time-series datasets. The proposed system brings a flexible, scalable, and intelligent solution to the realm of automated time series analysis using machine learning.

Introduction

The text presents AutoFeat Agent, an automated framework designed to improve time-series data analysis by solving key limitations of traditional feature engineering and machine learning approaches.

Modern time-series applications (finance, IoT, healthcare, climate, etc.) suffer from challenges such as manual feature engineering, temporal dependency issues, and data leakage (lookahead bias). To address these, the proposed system introduces a leakage-free, automated pipeline that performs feature extraction, selection, classification, and visualization.

The framework has four main layers: data preprocessing, automated feature engineering, machine learning evaluation, and an interactive dashboard. It uses tools like tsfresh for extracting thousands of statistical and temporal features from sliding time windows, ensuring only past data is used to prevent leakage. Feature selection is optimized using False Discovery Rate (FDR), reducing thousands of features to a smaller, meaningful subset.

For prediction, the system applies machine learning models such as Random Forest, Logistic Regression, and Gradient Boosting, with Gradient Boosting performing best. A Streamlit-based dashboard allows users to explore datasets, models, and results interactively.

Conclusion

This paper is about the AutoFeat Agent. It is a system that helps with time-series analytics. The AutoFeat Agent can automatically look at the data. Pick out the important parts. It also uses methods to make sure the data is good and useful. The system can even show the results in a way that\'s easy to understand. The people who made the AutoFeat Agent tried it out on some data. They found out that it works well and can make the data smaller and easier to use. The AutoFeat Agent is an flexible system that can be used for many different time-series analytics applications. In the future the people who made the AutoFeat Agent want to make it even better. They want to add some techniques like Transformer-based architectures and real-time streaming analytics. They also want to make it possible to use the system on datasets and to understand how it makes its predictions.

References

[1] M. Christ, N. Braun, J. Neuffer, and A. W. Kempa-Liehr, “Time Series FeatuRe Extraction on basis of Scalable Hypothesis tests (tsfresh – A Python package),” Neurocomputing, vol. 307, pp. 72–77, 2018. [2] F. Pedregosa et al., “Scikit-learn: Machine Learning in Python,” Journal of Machine Learning Research, vol. 12, pp. 2825–2830, 2011. [3] J. H. Friedman, “Greedy Function Approximation: A Gradient Boosting Machine,” Annals of Statistics, vol. 29, no. 5, pp. 1189–1232, 2001. [4] L. Breiman, “Random Forests,” Machine Learning, vol. 45, no. 1, pp. 5–32, 2001. [5] T. Chen and C. Guestrin, “XGBoost: A Scalable Tree Boosting System,” In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pp. 785–794, 2016. [6] Streamlit Inc., “Streamlit Documentation,” Available: https://streamlit.io/ [7] Python Software Foundation, “Python Language Reference,” Available: https://www. python.org/

Copyright

Copyright © 2026 Sakshi Bari, Rashmi Borse, Sarthak Patil, Meghana Patil, Manisha Patil. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.