RavenGraph provides graph‑derived signals and point‑in‑time embeddings for systematic trading, offering drop‑in features that are safe from look‑ahead bias. Its 64‑dimensional embeddings for over 300 US equities are updated daily and can be integrated offline via Parquet files, delivering higher Sharpe and predictive lift compared to traditional models.
Funding
Funding not disclosed
Founders
Product
Problem
Systematic traders often rely on handcrafted features or standard machine‑learning models that can suffer from look‑ahead bias and limited ability to capture inter‑asset relationships, leading to sub‑optimal risk‑adjusted returns.
Solution
RavenGraph supplies graph‑derived signals and point‑in‑time embeddings that encode the structure of asset networks for over 300 US equities. The 64‑dimensional embeddings are generated as of a specific timestamp, ensuring no future data leakage, and are refreshed daily (or intraday) for timely integration. Users can download the embeddings as Parquet files with a predefined schema, allowing offline evaluation and seamless incorporation into existing pipelines without live‑trading dependencies. Internal backtests show a Sharpe increase from 1.10 to 2.45 and a 15 % lift in predictive AUC compared with a baseline XGBoost model, indicating higher signal quality and regime robustness.
Target Audience
Primary customers are quantitative researchers and systematic trading teams that develop and backtest multi‑asset strategies, particularly those requiring bias‑free feature engineering.
Features
- 64‑dimensional graph embeddings for 300+ US equities, updated daily and intraday
- Point‑in‑time generation eliminates look‑ahead bias and normalization leakage
- Offline‑first delivery via Parquet files with schema for easy integration into research workflows
- Model‑agnostic signals that can be combined with any existing systematic trading strategy
- Built‑in signal ranking that highlights shock propagation pathways across assets
- Python client library for straightforward retrieval and merging with feature matrices