AutoScientist-Quant Automates Quantitative Investment Research

This paper introduces AutoScientist-Quant, a sophisticated framework of self-evolving coding agents designed to fully automate the quantitative investment research lifecycle. Current approaches to using large language model agents for discovering investment "alphas" often suffer from limitations such as an inability to adapt during execution, manual intervention required for library selection and model tuning, and potential lookahead bias in alpha discovery. AutoScientist-Quant addresses these by treating quantitative research as a single, budgeted search problem. A central controller dynamically makes decisions based on the remaining budget, determining whether to refine, combine, pivot, or halt the search, which nodes to expand, how many alphas to generate, and how to retrieve past trajectories from shared memory. This integrated core then handles library selection and model tuning, creating a closed loop from initial hypothesis to a deployable trading strategy. The framework also incorporates a robust evaluation pipeline, fixing two common lookahead problems and ensuring that the feedback window is distinct from the held-out test window, thereby guaranteeing true generalization. Experimental results across various CSI universes, backbones, and markets show that AutoScientist-Quant consistently achieves superior performance across nearly all metrics.