Intelligent Pre‑Loan Risk Control: Multi‑Loop Feedback Model, AutoML Controller Selection, Unsupervised Feature Extraction, and Metric Design

Akulaku, a Southeast Asian e‑commerce and fintech company, faces increasing pressure on its risk control system due to rapid growth in users, transactions, and business complexity. To address this, the company adopts two guiding principles: prioritize pre‑loan risk control and adjust it as promptly as possible.

Business Background

Founded in 2016, Akulaku reached 33 million registered users by 2020, leading to significant strain on its risk control infrastructure across pre‑loan, in‑loan, and post‑loan modules, each comprising credit flow, risk flow, and model flow.

Problem Analysis

The existing serial risk control process is modeled as a feedback (closed‑loop) system with four components: controlled object, controller, control algorithm, and feedback device. The system is abstracted into a linear ordinary differential equation to simplify analysis.

Practical constraints reveal that feedback only from post‑loan stages leads to delayed fraud detection and missed mitigation opportunities, prompting a shift to multi‑loop nested feedback control.

Solution Overview

1. Control Component Selection : The feedback control components map to risk‑control stages—controlled object (risk stage), controller (thresholds or switches), control algorithm (adjustment logic), and feedback device (output‑to‑input mapping). The solution emphasizes fault detection, localization, alert delivery, and recovery using AIOps techniques such as time‑series classification, anomaly detection, and root‑cause analysis.

2. AutoML‑Driven Controller Generation : AutoML trains fuzzy‑rule models that act as neural‑network controllers, fitting nonlinear differential equations while preserving interpretability. Data augmentation via a CNN‑based pipeline and Bayesian hyper‑parameter optimization mitigates small‑sample challenges.

3. Unsupervised Feature Extraction : Sequential ensemble techniques generate pseudo‑labels, while isolation‑forest depth serves as a proxy metric. Mutual information is approximated using Jensen‑Shannon divergence to detect noisy samples without labels.

4. Metric Quantification System : Input and output metrics are defined for each flow (amount, count, ratio). Control and feedback quantities are derived from model outputs and evaluated using filter, wrapper, and embedded methods (e.g., LASSO, tree‑based importance, deep k‑means).

Future Outlook

The authors propose a human‑machine optimal recommendation system based on metadata collection and controller deployment, and suggest moving from reactive post‑loss mitigation to proactive feedback‑driven control. They also discuss the limits of convex optimization in complex systems and advocate meta‑learning approaches to directly learn optimization algorithms.

In summary, the multi‑loop feedback control architecture, combined with AutoML, unsupervised feature extraction, and a robust metric system, aims to enhance the stability, steady‑state accuracy, and dynamic responsiveness of Akulaku's intelligent pre‑loan risk control.