Financial markets are noisy, non-stationary, and high-dimensional, making predictive and robust trading signals exceptionally difficult to discover. Alpha research has progressed from manual factor design to machine learning, evolutionary search, and LLM-based frameworks, improving how quickly researchers can generate, search, and evaluate candidate signals.
Yet most existing methods still automate isolated steps. They can produce factors or run evaluations, but they do not operate as end-to-end quant researchers that absorb external knowledge, connect a financial hypothesis to executable code, validate that implementation, and learn from accumulated discovery feedback.
From isolated generation to continuous research
XALPHA addresses this gap as a memory-driven AI quant researcher for continuous hypothesis-to-code alpha discovery. Its multi-source research memory combines financial knowledge grounded in external reports with empirical feedback from previous generations and research cycles.
Instead of treating each search as a fresh start, the system records what worked, what failed, and why. This accumulated research history shapes subsequent theme selection, hypothesis formation, implementation, and evaluation.
Macro, Micro, and Cross Brains
A Macro Brain plans research themes and selects suitable research Archetypes, using memory to determine which mechanisms and directions deserve further exploration. It then produces a structured pool of hypotheses rather than jumping directly into factor generation.
A Micro Brain transforms those hypotheses into executable factor code. Before evaluation, it verifies ex-ante tri-alignment among the original hypothesis, the implemented code logic, and financial plausibility, helping prevent a valid-looking program from silently testing a different idea.
A Cross Brain consolidates empirical outcomes into generation-level feedback, cycle-level summaries, and archetype-level research cues. Successful and failed experiments both become inputs to future exploration.
A closed loop that reads, validates, and evolves
Together, these components turn alpha mining from isolated factor generation into a closed-loop research process that continuously reads, hypothesizes, implements, validates, reflects, and evolves. The objective is not simply to create more candidates, but to make every research cycle better informed than the last.
Experiments on the CSI 300 show that XALPHA achieves stronger overall alpha-discovery performance than representative baselines. The results indicate that persistent research memory and explicit hypothesis-to-code validation can materially improve the quality of automated quantitative research.
The complete methodology, experiments, and author information are available in the paper on arXiv.