GEPS v5 — Graph-Guided Evolutionary Portfolio Search

Finance research idea generation and evaluation is automated through a graph-guided evolutionary portfolio search pipeline. It applies a seven-stage process including concept graph building, multi-channel generation, gating, Swiss ranking, verification, and portfolio optimization to produce a diversified set of top ideas.

Graph-guided evolutionary portfolio search is a structured, multi-stage pipeline that replaces debate-centric evaluation. It builds a concept graph from literature to seed diverse ideas, then applies gating, Swiss ranking, calibration, and portfolio optimization to generate a ranked set of research finalists.

To run the complete GEPS workflow, execute the /geps full command. This triggers the end-to-end pipeline, automating concept graph creation, idea generation, evaluation gates, Swiss ranking, verification, and portfolio optimization using your local literature data.

Yes, local literature data is required to use the GEPS v5 portfolio search. The system uses this data to build a concept graph that establishes relationships from literature, which then seeds the multi-channel idea generation and subsequent ranking stages.

Yes, the GEPS pipeline supports JSON-based inputs and outputs for integration. This allows you to interact with the modular Python tooling, including bradley_terry, calibration, and portfolio_optimizer components, to seamlessly incorporate the workflow into your existing systems.

A limitation of using a graph-guided concept graph for portfolio search is its reliance on local literature data quality. Additionally, while guardrails with taxonomy quotas and redundancy penalties exist, the structured pipeline requires modular Python tooling and JSON configuration to function correctly.