Automatically generate and stress-test quantitative trading factor ideas through a pipeline of specialized AI agents without manual research work.
Validate a proposed trading factor with statistical checks including information coefficient, Sharpe ratio, and drawdown before promoting it to a portfolio strategy.
Run isolated research sessions with a stated goal and inspect all experiment logs, checkpoints, and knowledge base files afterward.
Combine validated factors into a portfolio rotation strategy and backtest it out-of-sample against an equal-weight benchmark to check real-world viability.
Deep-learning factors require the extended PyTorch install, offline synthetic mode lets you test the full pipeline without live market data.
Robin is a research platform for quantitative finance that uses multiple AI agents working together to find, test, and evaluate trading factors and portfolio strategies. A trading factor is a measurable signal, such as price momentum or earnings growth, that might predict whether a stock will go up or down. Robin automates the process of generating ideas for these factors, stress-testing them, and deciding whether they are worth promoting to a portfolio strategy. The system runs a pipeline of specialized agents. One agent proposes factor hypotheses. Another writes three competing arguments about each factor: a bullish case, a bearish case, and a conservative one. A third agent implements the factor as either a mathematical formula or a small deep learning model. Then a validation agent runs statistical checks, measuring things like information coefficient, Sharpe ratio, and drawdown before deciding if the factor is good enough to keep. Accepted factors get combined into panels and then converted into portfolio rotation strategies, which get their own backtesting pass. The project keeps a clear boundary between research and trading: it places no actual trades and explicitly says it is not financial advice. On the technical side it separates what looks good in a backtest from what would qualify for production use. A strategy only advances if it beats an equal-weight benchmark out-of-sample, not just if it shows a strong Sharpe ratio in isolation. All runs produce auditable artifacts: experiment logs, checkpoints, knowledge base files tracking what was learned about each factor and strategy, and session transcripts. The session system lets you create isolated research runs with a stated goal, run them, and inspect the results after. Installing the package requires Python 3, and there are two install paths: the base install for traditional formula factors and an extended install that adds PyTorch support for deep-learning factors. A synthetic offline mode lets you test the pipeline without live market data. A daemon script is also included for running research on a recurring schedule.
← nenol2001 on gitmyhub — every repo by this author, as a profile.
Verify against the repo before relying on details.