FinWorld: An All-in-One Open-Source Platform for End-to-End Financial AI Research and Deployment
We introduce FinWorld, an open-source platform unifying data, training, backtesting, and deployment for financial AI research — covering time series forecasting, algorithmic trading, portfolio management, and LLM-based applications.
We introduce FinWorld, an all-in-one open-source platform for end-to-end financial AI research and deployment, accepted at KDD 2026.
Motivation
Financial AI research is fragmented. Data pipelines, model training, backtesting engines, and deployment infrastructure are typically built in isolation — making it difficult to reproduce results, compare methods fairly, or move from research to production. FinWorld solves this by providing a unified framework that covers the entire lifecycle.
What FinWorld Covers
FinWorld supports four core financial AI task types:
| Task Type | Description |
|---|---|
| Time Series Forecasting | Predict future asset prices and market indicators |
| Algorithmic Trading | Learn single-asset trading strategies via RL and LLMs |
| Portfolio Management | Optimize multi-asset allocation across customizable universes |
| LLM Applications | Financial reasoning, report generation, and agent-based analysis |
Key Features
Scale
- Over 800 million multimodal data samples spanning 1995–2025
- US and Chinese markets: DJ30, SP500, SSE50, HS300 indices
- Heterogeneous data: price, volume, news, filings, macroeconomic indicators
Architecture
FinWorld is built around a seven-layer modular design:
- Data Layer — unified ingestion and preprocessing for heterogeneous financial data
- Feature Layer — technical indicators, sentiment signals, and learned representations
- Environment Layer — realistic market simulation with transaction costs and slippage
- Model Layer — plug-and-play support for ML, DL, RL, and LLM-based models
- Training Layer — distributed training with experiment tracking
- Evaluation Layer — standardized backtesting and benchmark metrics
- Deployment Layer — live trading interface and monitoring
Benchmark Results
| Task | Model | Metric | Score |
|---|---|---|---|
| Time Series | TimeXer | MAE on DJ30 | 0.0529 (vs. LightGBM 0.1392) |
| Trading | SAC | Annual Return on TSLA | 101.55% |
| Portfolio | SAC | Annualized Return on SP500 | 31.2% |
| LLM | FinReasoner | FinQA / FinEval / ConvFinQA / CFLUE | State-of-the-art |
Links
- Paper: arXiv:2508.02292
- GitHub: DVampire/FinWorld
- Project Page: dvampire.github.io/FinWorld
