quantlab-factor-agent#

Version 0.2.0

quantlab-factor-agent mines daily stock factors with language models. A model states an economic mechanism, writes the factor it implies as KunQuant code, and a fixed pipeline decides whether the factor is new and adds to the factors you already have. The language model only proposes and reviews; every decision to accept is made by deterministic checks and statistics computed by quantlab on the training segment, which is all the model is ever shown.

The research loop of a campaign and the acceptance after it

Getting started

Install the package next to quantlab and run a first campaign from a config file.

Getting started
User guide

The research loop, the campaign config, the vocabulary, the prompts, the LLM backends, evaluation and acceptance.

User guide
API reference

Classes and functions: Campaign, CampaignConfig, Gate, Evaluator, Acceptance and the LLM backends.

API reference
Concepts

The terms the rest of the documentation uses: roles, mechanism types, the reference set, known factors, proposal groups, segments and what “accepted” means.

Concepts

What a campaign does#

            +-----------+   directions   +-----------+   candidates   +--------+
 ledger --> |  planner  | -------------> |  proposer | -------------> |  Gate  | -- refused: fix and retry
 memory     +-----------+                +-----------+                +--------+
                                                                          |
                    accepted <-- admission <-- Evaluator <-- reviewer <---+
                       |                       (training     (hypothesis,
                       +--> zoo, pool, ledger    segment)     code agree?)

Step

Who

Decides

Plan

planner (LLM)

one research direction per slot: a new mechanism, or a mutation or a refinement of an earlier trial

Propose

proposer (LLM)

a hypothesis, a description and KunQuant code for each direction

Gate

code

the code builds in a sandbox, reads only the vocabulary, is not a rewrite of an existing factor

Review

reviewer (LLM)

the hypothesis, description and code describe the same factor

Evaluate

code

IC, partial IC, correlations, deciles and coverage on the training segment

Admit

code

the factor clears the correlation, partial-IC t, monotonicity and coverage thresholds

After a campaign, acceptance trains a quantlab return model with and without the accepted factors on the validation segment, opens the hold-out once, and exports the chosen factors as quantlab factors and whole-market stores.

On this page