| ---
|
| license: mit
|
| tags: [benchmark, negative-result, time-series, binary-options, honest-odds]
|
| ---
|
|
|
| # NoEdge-Bench — Baseline
|
|
|
| Part of **[Nexus — The Honest Odds Project](https://nexusfinancial.sbs)**.
|
|
|
| A deliberately honest **null result.** The best simple model we can train to predict the next tick of a
|
| memoryless synthetic series, from strictly-causal features, scores:
|
|
|
| - **AUC = 0.5031**, accuracy = **0.5024** (permutation control AUC = 0.5004).
|
|
|
| That is chance level — **no better than a coin.** This is the whole point: there is **no player edge** on
|
| synthetic indices. The edge is in the payout, not the market.
|
|
|
| ### The look-ahead lesson
|
| An earlier version scored a fake ~0.74 because a centered rolling feature peeked at the future. Fixing it
|
| to **strictly causal** features collapsed it to ~0.50. If a "signal" only works with look-ahead, it is not
|
| a signal. This baseline exists so any submitted model can be measured against the coin.
|
|
|
| ## Use
|
| ```python
|
| from inference import predict
|
| p, _ = predict("noedge-bench-baseline", X) # p ~ 0.5 regardless
|
| ```
|
| safetensors weights + numpy inference (no torch). Not financial advice. License: MIT.
|
|
|
