gregyoung14/openmarket-models

OpenMarket Models

Pretrained model artifacts for the OpenMarket research archive. Model binaries are published here rather than committed to the OpenMarket GitHub repository.

How to Cite

@misc{openmarket_models_v02_2026,
  title        = {{OpenMarket} Binary-Outcome Calibration Model},
  author       = {Young, Gregory},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/gregyoung14/openmarket-models}},
  note         = {Release v0.2.1}
}

Repository Layout

v0.2.1/                            # current recommended release
  binary_outcome_model.json        # walk-forward logistic scorer (Platt scaling)
  binary_outcome_metrics_*.json    # training metrics + walk-forward windows
  model_manifest.json              # provenance manifest
v0.2/                              # prior full-unified training run
  ...
v0.1/                              # earlier release (smaller training set)
  binary_outcome_model.json
  binary_outcome_metrics_*.json
  model_manifest.json
README.md

Current Release (v0.2.1)

Artifact	Description
v0.2.1/binary_outcome_model.json	Calibrated logistic binary-outcome scorer
v0.2.1/binary_outcome_metrics_*.json	Walk-forward metrics (559 windows)
v0.2.1/model_manifest.json	Provenance manifest

Training pipeline (Rust):

1. export_step3_from_parquet on unified/ Parquet (v0.4.3-unified)
2. train_binary_outcome_model — walk-forward logistic regression + Platt scaling

Training data:

Field	Value
Dataset	gregyoung14/openmarket-btc-polymarket (v0.4.3-unified)
Feature export	step3_binary_calibration CSV from unified Parquet
Rows	357,390
Markets	2,251 / 4,450 in market_meta (51% write rate)
Notes	Backfilled unified/ synced to HF in v0.4.3-unified
Date range	2026-02-12 → 2026-05-14
Features	43

Aggregate metrics (calibrated, walk-forward OOS):

Metric	Value
AUC-ROC	0.838
Brier	0.165
ECE	0.025
Log loss	0.495

Simulated +EV trading (fee 1%, slippage 0.5%): 260,617 trades, 49.4% hit rate, -0.117 PnL/trade. Not deployable alpha — research artifact only.

Known limitations:

• Step3 export skips markets without sufficient Polymarket ticks or Binance trades (~50% of market_meta entries).
• Simulated economics are sensitive to fee/slippage assumptions.
• No ongoing model maintenance; frozen at source tag v0.5.2.

Reproduce:

cargo build -p step3-parquet-export -p binary-outcome-trainer --release
./target/release/export_step3_from_parquet \
  --parquet-root data/hf_release/unified_parquet \
  --out-dir data/hf_release/features_exports
./target/release/train_binary_outcome_model \
  --input data/hf_release/features_exports/step3_binary_calibration_<ts>.csv \
  --artifact-dir data/ml_artifacts

See scripts/ml/README.md in the source repo.

Previous Releases

v0.2

Same pipeline on pre-backfill unified Parquet (354,684 rows, 2,234 markets).

v0.1

Artifact	Description
v0.1/binary_outcome_model.json	Earlier calibrated logistic scorer
v0.1/binary_outcome_metrics_*.json	Metrics snapshots
v0.1/model_manifest.json	Provenance manifest

Trained on a smaller step3 export. Superseded by v0.2.1/ for research use.

Required Metadata Per Model

• source code commit
• dataset repo and dataset version
• feature schema (see feature_names in model JSON)
• training date
• validation split (walk-forward by market)
• metrics and calibration report
• known limitations

OpenMarket is in archival shutdown. Published artifacts are fixed research outputs, not an ongoing model release cadence.