Average accuracy (5s–1h)
50.7%
Across 8 short/mid horizons
Best horizon
5s
60.5% accuracy at best validation epoch
Significant edge (5s)
+8pp
vs. 52% baseline
Validation samples
~1.5M
Expanded walk-forward window
Performance by horizon

Accuracy at best validation epoch

Baseline (amber marker) is the 52% naive-direction benchmark used across all EventHorizon Crypto validation. Only the 5s horizon clears it by a statistically significant margin.

Horizon Accuracy vs. baseline (52%)
5s 60.5%
15s 52.4%
30s 48.0%
1min 50.0%
5min 50.2%
15min 49.8%
30min 50.5%
1h 50.5%
5h 54.0%
1d — no data
Statistically significant edge Within noise of baseline 52% baseline marker
Methodology applied

Validation pipeline

The 60.5% figure at 5s isn't a raw backtest number — it survived five separate checks before being published here.

01

Walk-forward, spaced sampling

Validation windows are spaced out in time, never adjacent to training data, to avoid temporal leakage.

02

Block bootstrap

Resampled with multiple block sizes to check that the confidence interval isn't an artifact of one block length.

03

Permutation test

Targets are shuffled and the model re-scored, to confirm the edge disappears when the real signal is destroyed.

04

Gap bootstrap

The gap between real and permuted performance is bootstrapped directly to test significance of the edge itself.

05

Temporal stability

Results are split into 15-day partitions to confirm the edge holds across time, not just in one lucky window.

Honest note

Real edge, unprofitable in practice

⚠️

The 5s edge is real — but exchange fees erase it

The +8pp edge at the 5-second horizon is statistically significant: it survives walk-forward validation, block bootstrap, permutation testing, and gap bootstrap. This is not noise.

It is also economically unviable for a retail trader under standard Binance fees. Net PnL is negative across every VIP fee tier we tested at this horizon — the edge is consumed by trading costs before it becomes profit.

We're publishing this as a complete case study: the validation proves the edge exists, and the fee analysis proves why it doesn't translate into profitable retail trading. Both halves matter.

Statistical edgeSignificant (+8pp)
Net PnL, all fee tiersNegative
Retail viabilityNot viable