Open-Source LLM Trading Experiment Seeks Contributors and GPU Compute
n1c39uy · reddit · 2026-08-27
The author is building an LLM-based trading research loop and is seeking contributors and compute. The project tests if an LLM can improve its own trading reasoning through iteration.
Mechanism:
- An LLM (currently Qwen2.5-27B) acts as a discretionary trader on historical 1-minute crypto data (multi-timeframe OHLC + indicators), outputting structured JSON trades (limit, stop, target, management rules) and reasoning traces.
- A simulation engine executes orders minute-by-minute with realistic fills and fees.
- Self-Improvement: Reasoning traces from winning trades are fine-tuned back into the model via QLoRA; the improved model generates the next round (rejection-sampling style).
Status & Needs:
- Running 4-bit inference and QLoRA on a single RTX 3090 (24GB) at 70 tok/s. Decision generation is slow (30-60s), and VRAM limits sequence length.
- Recruiting: Python/PyTorch programmers (backtest engine, GRPO trainer); 24GB+ GPU owners for batch inference; 48GB+ GPU owners for full-sequence fine-tuning.
- Performance: In a one-week backtest, the local 27B model (+19.1%) outperformed a frontier API model (+4.8%), though it breaks even in choppy markets.
More from Research
- Diverse pretraining reduces need for embodiment-specific data in robotics — chris_j_paxton · 2026-08-28
- PACT benchmark: one sentence of pressure raises AI rule violations 65%; no model clears unsupervised bar — baseten · 2026-08-28
- Two New Benchmarks Open-Sourced for Testing Agents in Dynamic Environments — AIwithGhotai · 2026-08-28
- Alex Rives, pioneer of protein language model ESM, named to TIME100 AI — proteinrosh · 2026-08-28
- New Paper: Dynamic Multi-Byte Prediction Speeds Up Hierarchical Byte-Level LMs — madeofAjala · 2026-08-28
- Jeff Dean's DiscoveryLoop Extends AI-Driven Iteration to Full Scientific Experiment Loops — agihouse_org · 2026-08-28