Agentic vs off-the-shelf LLM
The agentic pipeline delivers stronger ranking power and better compounding than zero-shot LLM factors; the process design drives the edge.
Agentic Factor Investing
We turn factor discovery into a repeatable engine. AI agents propose and combine candidate signals; a disciplined gate keeps only what works. What you see here is judged strictly on out-of-sample performance—no overfitting, no backtest bias.
Performance
Linear
Tree (LGBM)
Long-short cumulative return, out-of-sample. Hover for values; click legend to toggle series.
Key Out-of-Sample Results
Sharpe
Linear 3.82 / Tree 6.02
Long-short gross, out-of-sample.
Alpha after factor controls
The edge stays significant after market, size, value, profitability, investment, and momentum. Not explained away by standard factors.
Project Framework
Signals that rank returns
Daily scores separate future winners from losers. We aim for persistent cross-sectional edge, not one-off backtests.
Screen once, combine for scores
One gate for all candidates. Factors that pass enter the library; we then build composite scores with linear and tree models.
Out-of-sample first
We judge on data never used for selection. Performance shown gross and net of costs; edge tested against standard factor controls.
Summary and Outlook
The process creates the edge
Agent-driven discovery and combination beat naive LLM baselines. Gate, library, and combination design make the difference.
Cost-aware and factor-robust
The edge holds after trading costs and standard factor controls. We build for live conditions.
From prototype to platform
Scaling to a broader research platform: more automation, stronger risk controls, wider coverage.