The Digital Claustrum of Markets

The Contrarian Discovery

v0.2 of the MDL-DCC Predictor ran 7 generators (TREND-S, TREND-M, REVERT-MU, MOM-1, MOM-5, RW, VOL-REGIME) on 7,009 bars of BTCUSDT 15m data. Overall accuracy: 50.01% — dead coin flip. But the generator distribution revealed hidden structure:

REVERT-MU and TREND-M have marginal positive edge when they win. But TREND-S at 45.43% is actively harmful — worse than coin flip. The blended 50.01% is dragged down by bad generators that MDL still selects.

The Inverted DCC Spread

The DCC coupling parameter u was pegged at 1.0 in v0.2 (a bug: time-based updates never fired during a 30-minute backtest). After fixing to bar-based updates in v0.3, the DCC revealed its true signal:

The spread is -3.99% — inverted. When DCC says “I’m confident,” accuracy is worst. When DCC says “uncertain,” accuracy is near random. The DCC IS detecting real structure — it just correlates with generators being WRONG.

The MM Trap Detection Thesis

This inversion is not a bug. It is the central finding. Our generators ARE standard retail indicators: trend-following, mean-reversion, momentum. These are exactly what retail traders use. Market makers do the opposite.

When all standard indicators agree (high DCC u = compressible regime), that is when the MM trap is most loaded. Maximum retail consensus = maximum MM opportunity to reverse. The compressible pattern the DCC detects is not “price will continue this way” — it is “the MM has set a trap in this direction.”

The fix: contrarian mode. When u is high and generators agree strongly, flip the prediction. The 46.5% becomes 53.5%. The -3.99% spread becomes +3.99%. The architecture doesn’t change — only the interpretation.

Empirical Results: 924,481 Bars Across 12 Timeframes

The overnight analysis ran the DCC engine on BTCUSDT 1-minute data spanning June 2024 through March 2026 (21 months). The 1m data was resampled to 12 timeframes. Each timeframe was tested in both NORMAL mode (generators predict as-is) and CONTRARIAN mode (flip prediction when DCC says high confidence). 24 runs total.

All 24 Configurations, Ranked by Edge

(Table shows top 7 and bottom 4 of 24 runs. Intermediate rows omitted for clarity.)

Key finding 1: Zero normal-mode timeframes have positive edge. Every standard indicator loses on every timeframe over 21 months. The MM thesis holds: retail indicators predict the wrong direction.

Key finding 2: Contrarian mode produces positive edge on 8h (+0.68%), 15m (+0.54%), 30m (+0.52%), and 4h (+0.23%). The contrarian flip works, but only on medium-to-high timeframes where MM patterns have time to develop.

Key finding 3: VOL-REGIME dominates on all positive-edge timeframes. It detects volatility regime transitions — low vol (mean-revert) vs high vol (trend) — and the contrarian flip turns its predictions into correct MM-trap detection.

Best 2-TF Voting Combinations

When two contrarian timeframes agree on direction, the edge amplifies dramatically:

Key finding 4: The 1m arena alone has -1.22% edge (noise). But when 1m agrees with a higher timeframe, the combined signal is strongly positive. 1m appears in every winning combination. Cross-scale agreement amplifies signal that neither timeframe has alone.

Key finding 5: The 10m+1m combination produces +4.16% edge across 3,355 signals — the strongest finding. This is 5 signals per day over 21 months, a statistically meaningful sample.

Cross-TF Coherence

Pairwise direction agreement between contrarian timeframes is near 50% for most pairs (random), with notable exceptions:

Anti-correlated pairs (below 45%) are as interesting as correlated pairs — they indicate timeframes that systematically disagree, which could be used for regime classification.

The Fee Reality

The edge is real. The question is whether it survives execution costs.

The 10m+1m combo has +4.16% edge = 54.16% accuracy. On 10-minute BTC bars, the average absolute move is ~0.25%. The per-trade expected value is: 2 × 0.0416 × 0.0025 = 0.000208 = 0.0208% per trade.

With MEXC zero-fee perpetuals and leverage:

The Core Transfer: TSP ↔ Trading

In the TSP solver (8zrp v2.2–v2.4), DCC evolved from a single-knob intensity dial to a full governor controlling kick selection, intensity, restarts, worker coordination, and convergence detection. The trading DCC follows the same arc.

Markets have structure TSP lacks: multi-scale temporal organization. A 1h trend contains 4×15m swings, each containing 3×5m oscillations. The Multi-TF Governor (Chapter 9) exploits this. The contrarian flip (Chapter 1) exploits the MM adversarial structure that TSP also lacks.

DCC Sensing: The Ring Buffer Architecture

Implemented Specification (v0.3)

8-Bit Symbol Design

Bits 0-1: hit (11=correct, 00=wrong) — DOUBLE WEIGHT for LZ76
Bit 2:    regime changed (winner different from previous bar)
Bits 3-5: winner generator index (0-7)
Bit 6:    MDL margin bucket (0=weak, 1=strong)
Bit 7:    top-3 generators agree on direction

The double-weight hit bit is the key fix from v0.2. In v0.2, the winner-index bits (3 bits) dominated the LZ76 computation. When the same generator kept winning (stable = low LZ), u climbed to ceiling regardless of whether predictions were correct. In v0.3, the hit bits occupy 2 of 8 positions, ensuring LZ76 measures accuracy patterns, not just winner stability.

Actuator 1: Generator Governor

Thompson sampling with 0.997 exponential decay replaces winner-take-all MDL selection. Weak priors (0.5/0.5) ensure first observations dominate. Bad generators lose weight automatically — TREND-S at 45% accuracy decays to near-zero Thompson score within 200 bars.

The blend temperature is controlled by u: high u (stable) = winner-take-all (temp=0.5), low u (chaotic) = broad blend (temp=5.0). When generators disagree on direction, the blend produces near-zero confidence — which maps to “reduce size” or “don’t trade.”

In contrarian mode, the blended direction is flipped when u ≥ 0.6 and confidence ≥ 0.3. This is the mechanism that converts the -3.99% inverted spread into positive edge.

Actuator 2: Window Governor

DCC controls the MDL window dynamically. High u = expand to 2× base (400 bars at base=200) for better statistical power. Low u = shrink to base/4 (50 bars) for fast regime adaptation. Changes are momentum-damped: max 10% per update. Overnight data shows actual window range was 50–400 with mean ~310.

Actuator 3: Regime Reset Governor

Five-level escalation ladder, bar-based (critical fix from v0.3 — time-based never fired during fast backtests):

Overnight runs triggered 2 nuclear resets on 15m BTC across 6,808 bars, confirming the escalation fires on real data at appropriate intervals.

Actuator 4: Multi-Timeframe Governor

This is where trading DCC goes beyond TSP. The overnight results validate the architecture empirically.

The 1m arena alone has -1.22% edge (noise). The 10m arena alone has -0.22%. But when both agree in contrarian mode, the combined edge is +4.16% across 3,355 signals. Neither timeframe has meaningful edge alone. The cross-scale agreement creates edge from nothing.

This validates the core Multi-TF Governor thesis: run arenas at multiple scales, and the governor’s job is to detect when they agree. Agreement across scales means the structure is real (visible from multiple vantage points), not noise (visible from only one).

Actuator 5: Position Governor

DCC coupling u and multi-TF coherence together control sizing:

The Market S-Metric

S = coherence × complexity, from the Claustrum-Consciousness Hypothesis.

Coherence = cross-TF direction agreement. Overnight data shows: 15m+1D = 60.4% (high), 3m+8h = 39.3% (anti-correlated). Coherence varies by pair, creating a real-time signal.

Complexity = Shannon entropy of which generators win across TFs. If VOL-REGIME wins everywhere (as it did on positive-edge TFs), complexity is low. If different generators win on different TFs, complexity is high.

In v0.3, S is logged but not controlling decisions. The overnight data provides the first real measurements to correlate with profitability. If high-S periods (coherent direction + diverse generators) correlate with higher combo accuracy, S becomes a control signal in v0.4.

The Search Space Problem

Finding the optimal DCC configuration across all TF × asset × combo permutations IS a combinatorial optimization problem — and we already have the tools to solve it.

L1 → L2 → L3: The Unified Optimizer Pattern

For seconds-scale: use the 1s BTC data (86,400 bars/day). Resample to 2s, 3s, 5s, 10s, 15s, 30s. Run same engine. MM stop hunts are 1-second events — the DCC might see them as highly compressible because they follow a template.

For cross-asset: BTC 10m + ETH 5m. MMs move correlated assets together — a BTC dump triggers ETH/SOL dumps 2-5 seconds later. Cross-asset DCC coherence detects the MM propagation delay.

Meta-DCC: DCC Governing the Search Itself

The search for the best TF combo IS a compression problem. Each combo configuration is a “generator.” MDL scores how well each generates profitability. DCC governs the search budget: when finding new edges (low complexity = compressible search), explore more. When stuck (high complexity), stop and zoom elsewhere. This is TSP v2.4’s DCCGovernor applied recursively to the optimizer itself.

The Three-Engine Architecture

DCC / SM / ZZ: Clean Responsibility Split

In the hybrid: ZZ makes the trading decision. DCC tells ZZ “the regime is contrarian-favorable on 15m and 30m” or “regime is ambiguous, reduce size.” SM tells ZZ “volume confirms this breakout” or “absorption detected, this is a trap.” Three independent lenses on the same market. No duplication.

Volume Ownership: SM, Not DCC

SM already has complete volume-price analysis: absorption detection (big vol + small candle → reversal), momentum confirmation (big vol + big candle → continuation), and volume divergence (price up + volume down → weak). DCC uses price structure only. This is deliberate: DCC’s power comes from compression of price dynamics. Adding volume to DCC would duplicate SM’s work without adding new signal.

The one exception: a future VOL-ABSORB generator that detects the absorption pattern and enters the MDL arena. This adds volume signal to DCC’s direction prediction without duplicating SM’s trade-gating logic. Planned for P2.

Signal Quality Monitor

Tracks rolling accuracy across all TFs. Output: ACTIVE or DARK.

active = (best_tf_accuracy > 0.505
          and best_tf_u > 0.4
          and escalation_level < WITHDRAW)

DARK mode: DCC still runs, still measures, still logs. But output to trading system is “no signal.” Capital preservation. The system re-engages automatically when structure returns.

Full Auto Mode

# Current (v0.3):
python BD_MDL_DCC_Predictor.py --file data/BTCUSDT_15m.csv --window 200

# Full auto (target):
python BD_MDL_DCC_Governor.py auto BTCUSDT --max-hours 24

In auto mode, the DCCGovernor runs: data fetch → resample all TFs → discovery phase (generator calibration) → warmup (DCC calibration) → active trading (all actuators live) → report. The human says “here’s the asset, here’s the time.” DCC decides everything else.

Implementation Phases

Connection to 8Z-OS

The DCC Trading Governor is the eighth domain where MDL + DCC produces results — and the first with empirical validation on 21 months of real market data.