mean-reversion-24h

Pre-registered evaluation report · generated 2026-08-18T18:45:19.747832+00:00 by nullbar 0.6.0

1The frozen registration

The sha256 above is of exactly the text below, with no trailing newline. Re-hash it to confirm the design and the bar are the ones that were frozen.

{
  "bar": {
    "gross_2x_cost": {
      "metric": "gross",
      "op": ">=",
      "value": 0.46
    },
    "null_flat": {
      "metric": "null_max_abs_t",
      "op": "<",
      "value": 3.0
    },
    "t3": {
      "metric": "t",
      "op": ">=",
      "value": 3.0
    },
    "year_2022_positive": {
      "metric": "y2022",
      "op": ">",
      "value": 0.0
    }
  },
  "cells_budget": 64,
  "created_at": "2026-08-18T18:44:54.484739+00:00",
  "design": {
    "cost_pct_rt": 0.23,
    "hold_bars": 24,
    "oos": "excl 2024-10-01..2026-01-01",
    "signal": "dist_ma168 <= expanding shifted q10",
    "universe": [
      "BTC-EUR",
      "ETH-EUR",
      "SOL-EUR",
      "XRP-EUR",
      "ADA-EUR",
      "LINK-EUR",
      "LTC-EUR",
      "AVAX-EUR"
    ]
  },
  "hypothesis": "bottom-decile dist_ma168 vs own history mean-reverts over 24h on the 8-symbol EUR universe",
  "name": "mean-reversion-24h"
}

2The search

Deflation counts cells the way |t| does: a signal tested long and short is one two-sided cell, not two. Cells used below: recorded on the test look.

3The null control

A machinery check, not an edge: it asks whether the pipeline invents an effect on scrambled returns beyond what the assets it holds pay unconditionally.

4The held-out result

5Fill realism

Under resting-fill pricing this record has no gross left to haircut: the touch leg is negative. The ratio above crosses zero and is not a fraction that survived.

The entries that never fill are disproportionately the good ones, so the assumed row is an upper bound on what any of this could have executed at.

6Deflation

The record claims 32 cells against 64 ledger rows (two-sided pairing counts one cell per signal tested in both directions); the threshold below is the smaller, weaker deflation. Thresholds are simulated at report time from the recorded cell and cluster counts (100,000 draws, seed 0) — seeded, so they reproduce. Sharpe: recorded. Deflated Sharpe: recorded. The MEAN of the noise maximum is not a bar — pure noise beats its own expected maximum about 45% of the time. The 5% line is.

7The bar, as graded

Graded against the file on disk. Unregistered conditions are ignored: adding one until something passes is the failure this library exists to prevent.

8Anchoring

Local only. No remote branch contains these commits, so nothing outside this machine has seen them and the whole history could be rebuilt in a minute.

What a git anchor proves: that the bytes committed are the bytes graded, that the registration commit precedes the test-look commit, and that both are still reachable from HEAD. What it does NOT prove: wall-clock time — commit dates are self-reported and one environment variable forges them, so only a push to a host the researcher does not control was witnessed by anyone else. It also cannot show that the researcher had not already seen the test window: ordering of documents is not ordering of knowledge.

9The rest of the record

Every remaining top-level entry of the test-look payload, so nothing recorded is hidden by this rendering.