Measurement window: 29 July – 13 August 2026 · Last updated: 7 September 2026
← bitdataview.com · Türkçe oku →
Analytics platforms usually tell you only what they found. We publish what we tried — whatever the outcome. This page is an open record of the candidates we have tested in search of an indicator that moves before price, the measurements they produced, and the reasoning behind each decision.
Every candidate so far has been rejected. We are not hiding that, because a candidate tested and left unrecorded gets retried from scratch months later. This page is also our own memory — and when a measure does clear the gates, it will be written up here in the same detail.
Everything we measured appeared after the price move, or at the same time as it. Nothing appeared before.
Nineteen separate candidates, nineteen different methods, one result. These are not nineteen individual failures — they are a systematic finding. What that means, and why they all failed at the same point, is covered in the sections that follow.
The same discipline will apply when a measure does work: the same gates, the same detail, the same openness.
None of these measurements changed the production signal path. A candidate rejected in research never entered the signals you see. For a finding to reach the live system it must clear every criterion written down in advance; none of these candidates did.
Each candidate was asked a single question: does this measure move before price does? The table below gives the candidate and the point at which it failed. Where the name is a link, a full write-up is available — how we measured it, what numbers came out, and why it was rejected. We are adding the rest one at a time.
| # | Candidate | How it failed |
|---|---|---|
| 1 | Price derivatives (RSI / ATR / wavelet / MF-DFA) | Produced no independent contribution |
| 2 | Volume profile / support strength | Tautology — with volatility held constant the effect fell from +0.103 to +0.0008 |
| 3 | Trade count measures | Weak — even the strongest was half the size of the reference measure |
| 4 | CVD slope | Contamination, a suppressor effect, and inconsistency in 7 of 10 symbols |
| 5 | CVD divergence (four-cell) | Did not hold in a disjoint window or across two separate bull periods |
| 6 | CVD conditioned on regime | The sign reversed in 9 of 12 cells |
| 7 | CVD regime transition | Only 9 transitions; the bottom candidate held in 4 of those 9 |
| 8 | Volume regime gate | The sign came out fully inverted in a disjoint window |
| 9 | Volume node + taker flow | The sign reversed once measured directionally |
| 10 | Volume / ATR filter | No measurable lift in a clean window |
| 11 | Open interest divergence | Effect near zero; present in 4 of 8 symbols |
| 12 | Stablecoin supply | Simultaneous — the lag curve peaks exactly at zero |
| 13 | Cross-asset relationships (Gold / DXY / S&P 500) | All signal sits at zero lag; no forward-looking information |
| 14 | Mandelbrot → Lorenz approach | 51.0% accuracy over 991 bars — indistinguishable from a coin flip |
| 15 | Wick asymmetry (candle geometry) | Held in 4 of 18 symbols; sign unstable across the lag curve |
| 16 | Volatility change (d(vol)/dt) | Genuinely distinct from the level — and equally directionless |
| 17 | Drawdown recovery speed | Lag curve peaks at zero (a mechanical link); the sign reversed within regimes |
| 18 | ATR → regime change | Sign unstable; event count off by a factor of 16 from expectation |
| 19 | Acceleration (second difference of four measures) | All four failed at the same gate; differentiating also destroyed existing structure |
Three further attempts were rejected alongside these: post-buy-signal filter effects (statistically indistinguishable from nothing), a sell-signal relationship with the Asian session (the effect evaporated as the sample grew), and volatility clustering (confirmed as a real phenomenon, but it carries no directional information).
On 13 August, five candidates were measured on the same day. Expectations were written down before the data was seen, and all five were rejected. No selection was made among them; because every one failed, the question of "did we pick the best-looking result" never arose.
Three separate conclusions came out of that day:
Why publish this? Saying "there's a correlation" is easy. The hard part is measuring whether that correlation forms before price or at the same time. Most candidates fail exactly here — and a system that doesn't draw that distinction will look excellent in hindsight while telling you nothing about what comes next.