1. The Market Question
Between November 4 and November 5, 2026, Bitcoin executed a textbook $80,000 round-trip:
- • Act I (Nov 4 Breakout): BTC rallied from
$75,658(03:09 UTC) through$80,000, peaking at$81,042(15:18 UTC). - • Act II (Nov 5 Plateau & Breakdown): For 6.5 hours on the morning of Nov 5 (
05:00–11:30 UTC), BTC held a tight, seemingly stable plateau between$80,156and$80,700(median$80,351). Then, after11:30 UTC, the bid vanished and spot dropped -6.1% to$75,449.
To a trader watching spot price or perpetual funding alone, the Nov 4 breakout above $80,000 and the Nov 5 morning consolidation above $80,000 looked structurally identical.
They were not. Under the surface, cross-market conviction in prediction markets (Q2) and knowledge-graph capital flow (Q3) had already inverted completely.
2. Key Findings at a Glance (Zero Timestamp Cherry-Picking)
Instead of selecting a handful of favorable timestamps after the fact, we audited 100% of the out-of-sample production records generated across the 42-hour window (N = 2,727 Sybil-clean 16-channel feature snapshots and N = 3,263 multi-horizon price forecasts):
| Audit Dimension | Empirical Finding Across Full 42-Hour Population |
|---|---|
1. Regime Separation (Act I vs. Act II) | In the 6.5 hours before the Nov 4 breakout (n = 426 ticks), informed prediction-market consensus (Ch 5) was bullish (> 0.50) on 81.9% of snapshots (mean 0.564, peaking at 0.875). In the 6.5 hours before the Nov 5 breakdown (n = 462 ticks), while BTC sat at $80,350, it was bullish on 0.0% of snapshots (0 out of 462; mean 0.353). |
2. Structural Feature Rank IC (4h–8h) | Raw Q2 (Prediction Market) and Q3 (Causal Graph) tensor features achieved tick-level Spearman Rank IC of +0.38 to +0.55 at 4h–8h horizons, and +0.37 to +0.48 (p = 0.003 to 0.027) when subsampled onto a 1-hour (60m) grid (n = 33–35). |
3. Where the Model Failed (point_forecast at 5m–4h) | Short-horizon price-level forecasts (5m, 15m, 4h) had near-zero or negative Rank IC (-0.057 to -0.021) and underperformed a Random Walk in RMSE. Even at 24h, spot-momentum drag caused false-positive bullish price projections on Nov 5 while the underlying Q2 feature channels remained 100% bearish. |
| 4. Recommended Deployment for Systematic Desks | Do not trade the short-horizon price regression head. Consume the raw Point-in-Time Q2 and Q3 feature states (informed_market_prob, entropy_conviction_psi, informed_capital_imbalance, financial_impact_flow) directly as 4h–8h structural regime filters and position-sizing overlays. |
3. The Evidence: Act I ($78.2k -> $81.0k) vs. Act II ($80.3k -> $75.4k)
Table 1 compares the entire population of Sybil-clean snapshots in the 6.5-hour pre-breakout window on Nov 4 (n = 426) against the 6.5-hour pre-breakdown plateau on Nov 5 (n = 462).
Table 1: Full-Window Regime Comparison (n = 426 vs. n = 462 Snapshots)
| Signal Channel | What It Measures | Act I: Pre-Breakout (Nov 4, 08:00–14:30 UTC, n = 426) | Act II: $80.3k Plateau (Nov 5, 05:00–11:30 UTC, n = 462) | Structural Takeaway |
|---|---|---|---|---|
informed_market_prob (Q2, Ch 5) | Reputation-weighted probability across active macro/crypto prediction markets ([0, 1], 0.50 = neutral) | Mean: 0.564 ± 0.100 81.9% of ticks > 0.50 (349/426)(Peaked at 0.875 during breakout) | Mean: 0.353 ± 0.037 0.0% of ticks > 0.50 (0/462)(100% of ticks below 0.50) | Smart prediction-market capital completely abandoned the $80k plateau hours before spot broke. |
entropy_conviction_psi (Q2, Ch 6) | Liquidity-adjusted conviction strength ([-1, +1], penalizes thin orderbooks) | Mean: +0.162 ± 0.207 84.5% positive ( 360/426) | Mean: -0.335 ± 0.098 98.5% negative ( 455/462) | High-depth conviction flipped from strongly positive to persistently negative. |
informed_capital_imbalance (Q2, Ch 8) | Net orderflow divergence between top-tier wallets and retail ([-3, +3]) | Mean: +0.269 ± 0.626 78.9% positive ( 336/426) | Mean: -0.715 ± 0.265 98.5% negative ( 455/462) | Tier-1 capital was net-accumulating YES on Nov 4 and net-distributing into retail on Nov 5. |
financial_impact_flow (Q3, Ch 10) | Decaying authority-weighted capital flow from verified graph events (>= 0) | Mean: 0.576 ± 0.117 76.5% of ticks > 0.50 | Mean: 0.453 ± 0.070 21.4% of ticks > 0.50 (-21.4% decay) | Institutional inflow catalysts decayed and were replaced by miner treasury sales and ETF outflows. |
derivatives_surface_state (Q1, Ch 4) | Perp funding rate & skew composite ([-1, +1]) | Mean: +0.028 ± 0.216 54.5% positive | Mean: +0.063 ± 0.348 47.6% positive | Blindspot: Perp derivatives looked nearly identical in both windows, giving no early warning on Nov 5. |
If a desk relied on Q1 perpetual derivatives (derivatives_surface_state), Nov 4 (mean +0.028) and Nov 5 (mean +0.063) were indistinguishable. By contrast, even if we collapse the minute-level data into 13 strictly non-overlapping 1-hour block means (6.5h in Act I vs. 6.5h in Act II), Q2 (informed_market_prob) separated the two regimes with 100% accuracy and zero overlap (p = 0.0012 across N_eff = 13 independent hourly blocks).
4. Signal Predictive Power vs. Forecast Head Leakage
A core architectural finding of this audit is the separation between feature information content and regression head execution:
Table 2: Feature Rank IC vs. Forward Realized BTC Returns (4h–8h Horizons)
| Feature Channel | 4h Tick IC (n=2,362) | 6h Tick IC (n=2,203) | 8h Tick IC (n=2,018) | 6h Hourly-Grid IC (n=35, p_60m) | 8h Hourly-Grid IC (n=33, p_60m) |
|---|---|---|---|---|---|
Ch 6 entropy_conviction_psi (Q2) | +0.406 | +0.549 | +0.549 | +0.483 (p = 0.0033) | +0.484 (p = 0.0044) |
Ch 5 informed_market_prob (Q2) | +0.381 | +0.534 | +0.521 | +0.406 (p = 0.0156) | +0.418 (p = 0.0155) |
Ch 8 informed_capital_imbalance (Q2) | +0.334 | +0.507 | +0.503 | +0.385 (p = 0.0224) | +0.389 (p = 0.0251) |
Ch 10 financial_impact_flow (Q3) | +0.463 | +0.487 | +0.396 | +0.374 (p = 0.0270) | +0.309 (p = 0.0797) |
Econometric Caveat on Sample Size (N_eff): Because a 42-hour event study contains one major upward impulse and one major downward impulse (N_eff = 5to10strictly disjoint4h–8hwindows), single-event RankICs (+0.38to+0.55) reflect regime-separation fidelity during a macro inflection, not a claim of+0.50steady-state dailyICacross all market environments. Multi-month walk-forward benchmarks across choppy, trend, and macro-catalyst regimes are maintained in the full validation suite.
Why Raw Features Beat the point_forecast Regression Layer
When we audited all N = 3,263 price forecasts (forecast_accuracy_emissions) across all 7 horizons (5m to 24h):
- Short Horizons (
5m,15m,4h) Leak Alpha: The price-level regression head anchors heavily to recent spot price (Q1). At5m(IC = -0.021),15m(IC = -0.043), and4h(IC = -0.057), the price forecast head fails to beat a zero-drift Random Walk. - Long Horizons (
8h&24h) Capture Tail Moves, But Suffer From Momentum Drag:- • At
8h(480m), large predicted dislocations (>= 200 bps,n = 71) achieved64.8%directional accuracy, with71.2%bullish precision (52 TP / 21 FP). - • At
24h(1440m), bearish dislocations (<= -100 bps) achieved96.7%precision (29 TP / 1 FP)—calling the$80,327 -> $75,603breakdown hours in advance. However, while spot sat near$80,350on Nov 5,Q1price momentum dragged the24hregression head into emitting 21 false-positive bullish forecasts alongside its 14 true-positive bearish calls. - • Meanwhile, the raw
Q2feature channels (Table 1) emitted0false-positive bullish ticks (0/462) across that entire 6.5-hour window.
- • At
5. Practical Takeaway for Quantitative PMs & Researchers
The value of cross-market causal infrastructure is not in predicting the next 5-minute candle on Binance. It is in detecting when spot price is diverging from informed capital consensus.
- • How to Use This Signal: Systematic desks and risk managers should ingest the Point-in-Time
Q2andQ3tensor channels (informed_market_prob,entropy_conviction_psi,informed_capital_imbalance,financial_impact_flow) directly as 4h–8h regime filters, conviction-sizing multipliers, or drawdown throttles—bypassing the short-horizon price regression head. - • Want Us to Audit a Move on Your Book? Give us a ticker and a timestamp window where a move caught your desk or models off-guard. We will pull the Point-in-Time 16-channel tensor and run the exact out-of-sample regime & attribution breakdown.
For the complete 16-channel mathematical specification, confusion matrices across all 7 horizons, disjoint-block autocorrelation tables, and row-level timestamp logs, see the companion .