no-edge-detector

No-Edge Detector

Part of Nexus — The Honest Odds Project.

Given a price series, this model classifies it as:

• memoryless_no_edge — statistically indistinguishable from a random walk (e.g. Deriv synthetic indices);
• mean_reverting — shows mean-reversion structure (e.g. real-underlying baskets);
• trending — shows momentum/trend structure.

It detects statistical structure, not profit. Structure is necessary but not sufficient for an edge — it still has to survive spread/payout. On Deriv synthetics there is no structure at all; on real-underlying baskets there is mean-reversion, but the payout eats it. This is not financial advice and predicts no prices.

Features (all causal)

acf1, acf2, acf3, acf4, acf5, absacf1, absacf2, absacf3, vr2, vr4, vr8, hurst, runs_z, er — autocorrelation (returns + |returns|), variance ratios, Hurst exponent, runs-test z, efficiency ratio; computed on a window of 256 points.

Metrics

Held-out accuracy: 0.940 (3-class). See metrics.json.

Use

import numpy as np, common as C
from inference import predict
probs, labels = predict("no-edge-detector", C.features_from_prices(my_price_series))

safetensors weights + numpy inference (no torch). License: MIT.