Title: AlphaEval: A Unified and Efficient Framework for Assessing Formula Alpha Mining

Formula alpha mining, a process essential to quantitative investing that derives predictive signals from financial datasets, has seen significant advancements through algorithmic methods like genetic programming, reinforcement learning, and large language models. Despite these gains in discovery capacity, the field still faces substantial hurdles in systematic evaluation. Current assessment standards largely rely on backtesting and correlation-based metrics, both of which have notable limitations. Backtesting is notoriously computationally demanding, operates sequentially, and exhibits high sensitivity to specific strategy parameters. Conversely, while correlation-based metrics offer computational efficiency, they focus narrowly on predictive capability, neglecting other vital attributes such as temporal stability, robustness, diversity, and interpretability. Furthermore, the proprietary nature of most existing alpha mining models impedes reproducibility and slows broader scientific progress.

To resolve these challenges, we introduce AlphaEval, a novel evaluation framework designed for automated alpha mining models. AlphaEval is unified, parallelizable, and eliminates the need for backtesting. It evaluates the holistic quality of generated alphas across five distinct, complementary dimensions: predictive power, stability, robustness against market perturbations, financial logic, and diversity. Our extensive experiments, conducted across various representative alpha mining algorithms, reveal that AlphaEval delivers evaluation consistency on par with comprehensive backtesting, while offering deeper insights and superior efficiency. Moreover, the framework proves more effective than traditional single-metric screening methods in identifying high-performing alphas. In an effort to foster reproducibility and encourage community involvement, we have open-sourced all implementations and evaluation tools associated with AlphaEval.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC