failed-market-maker

The Failed Market Maker

An archive of a market-making project that did not establish a learned, independently verified live trading edge. It includes the weights, the promising results, the losses, and the evidence that changed the interpretation of those results.

完整中文 · Models · Results · Development story · Inference · Provenance and redaction

GitHub · Hugging Face · GitHub model release

The project explored Polymarket BTC UP/DOWN five-minute markets using Polymarket order books, Binance market data and Chainlink observations. It evolved from a replay engine and feature pipeline, through supervised signals and offline policies, into an Alpha + MakerPolicy deployment and a newer architecture. This release preserves that progression without rebuilding the training or trading system.

What is here

There are 24 retained trained candidates: 3 M3, 1 M5 behavior-cloning model, 3 M5 IQL models, 8 Alpha models, 7 MakerPolicy models and 2 Phase 6.7 repair/control candidates. The Alpha + Policy live deployment is listed separately; adapted exports and deployment copies are not extra trained candidates. The other eight Phase 5.8 primary experiments remain in the experiment ledger even though their weights were not retained.

The June 2 paper run reported +1,791.3144 USDC adjusted PnL for the R1 Alpha + Policy combination, but decisions stopped on June 5. It was not a completed seven-day test. The June 22 12:06 live report recorded −18.87545541 USDC local conservative NAV change, including unrealized inventory; it was not a final platform account reconciliation. Both belong in the record. Run-specific evidence

Download and run

The repository contains the bilingual narrative, model cards, contracts, original metric reports and inference code. GitHub Release holds a model archive with the same models/<model_id>/ directory structure used on Hugging Face. Extract the archive into this repository root, or obtain the complete Hugging Face repository. Preserve companion files: the BC ONNX requires its .onnx.data file, and contracts/normalization must travel with each model.

python -m inference.run models/m3_run3/best_model.pt --family m3 --smoke
python -m inference.run models/m5_v7v72_bc/agent_model_m5.onnx --smoke
python -m inference.run models/mpv2_r1_a3_seq120_seed2 --smoke

Use an environment with the dependencies in requirements.txt. Inference instructions describe actual input ordering and the distinction between a constructed-input check and historical evaluation. These commands place no orders and do not train or replay markets.

Reading the archive

manifest.json is the unified inventory: model identities, exact files, byte counts, logical source locations, transformations, licenses and completeness gaps. MODELS.md links every bilingual model card. RESULTS.md separates training, offline replay, paper and live results. Missing values always include a reason; they are never replaced with zero.

The newer architecture's formal trained model files, training scores and paper/live profitability are N/A — no formal trained release in this archive. Its component and interface checks are engineering evidence. Historical documents under evidence/ are source records, not current deployment instructions; their relative links may refer to private historical files not included here.

Original code and weights are licensed under Apache-2.0. Curated documents and reports are licensed under CC-BY-4.0. Attribution and dependency boundaries appear in NOTICE. Raw market datasets are not distributed.