01
Research contextContext
Review previous experiments and failures to ground the next idea.Use previous experiments to guide new research.
Automatic Quant Research System
One agentHypothesize · interpret
Why we built KwontFactory
We wanted to reduce the bottleneck of human-led hypothesis development, coding, and interpretation in quant research. KwontFactory connects an LLM agent doing that work with tools that execute its experiments reliably.
The research bottleneck
In a conventional quant workflow, a researcher develops an economic hypothesis, chooses data and trading rules, and writes the code. After a backtest, they interpret the result and revise the hypothesis or implementation. Even when calculations are automated, deciding what to test and how to revise it still requires human time and judgment.
This repeated work is the bottleneck KwontFactory addresses. The starting point was to use an agent’s reasoning and coding abilities to carry research beyond the pace at which one person can implement and review each experiment.
The agent’s role
An LLM agent brings the ability to read context, propose hypotheses, express them in code, and revise its work from feedback. Its role is to explain how a market mechanism might create a trading opportunity and turn that idea into a testable strategy.
The agent reads the result: did the expected behavior appear, did costs erase returns, or was there an implementation problem? That interpretation guides the next hypothesis and revision.
Purpose-built evaluation tools
Backtesting and statistical tests, however, must produce the same result from the same inputs and conditions. Having an agent rebuild those calculations for every experiment, or judge their validity through prose, adds work and can make evaluation inconsistent.
Strategy compilation, cost-aware backtesting, statistical tests, and experiment recording are therefore implemented in advance as deterministic Python tools. The agent invokes them and reads their results. It can revise a hypothesis while the tools apply consistent calculation and validation rules.
A harness for the research loop
A collection of tools alone does not carry research forward. Their inputs, outputs, records, and workflow must let the agent review earlier experiments, submit a new strategy, and use the evaluation in its next decision. The harness provides that research environment.
KwontFactory is designed around one agent repeating a loop of hypothesis, strategy authoring, tool-based evaluation, and feedback. Agent intelligence drives exploration and interpretation; prepared tools handle repeated calculations and validation.