| --- |
| license: mit |
| library_name: sklearn |
| pipeline_tag: tabular-classification |
| tags: [polymarket, prediction-markets, the-oracle] |
| --- |
| |
| # THE ORACLE — direction model (logreg.updown.v1) |
|
|
| Predicts whether a Polymarket market's YES price will be **higher 7 days |
| from now**. Trained by THE ORACLE, an autonomous agent funded by $ORACLE pump.fun |
| creator fees. |
|
|
| ## Honest backtest (held-out, later-in-time, no look-ahead) |
| | metric | value | |
| |---|---| |
| | ROC-AUC | **0.65507** (0.5 = no skill) | |
| | accuracy | 0.67072 | |
| | directional accuracy (on moves) | 0.6018 | |
| | up-rate (class balance) | 0.33114 | |
| | samples (train / test) | 23661 / 5916 | |
|
|
| A companion regression confirmed that move *magnitude* cannot beat the market's own |
| price (skill ~0). The edge is in **direction**, which is what this model captures. |
|
|
| ## Data & features |
| Real daily price curves of active Polymarket markets (CLOB `prices-history`, |
| `interval=max&fidelity=1440`). Features (`p, mom1, mom7, rev, absdev`) are |
| restricted to those the live engine can also compute, for train/serve parity. |
|
|
| ## Scoring (pure JS, no deps) |
| `std = (x - mean)/scale; pUp = sigmoid(Σ coef·std + intercept)` |
|
|
| Not financial advice. |
|
|
