A Causal Approach to Bitcoin Performance Modeling

01 — Context

Research Question

Bitcoin has attracted significant institutional interest, yet a fundamental question remains unresolved.

🔗 The Problem

Traditional correlation-based analysis is unreliable: correlations between Bitcoin and other assets are unstable over time and highly sensitive to the investment horizon and macro-financial context. Should Bitcoin be in a core portfolio for diversification, or confined to a satellite allocation?

🎯 Our Approach

We shift from associational analysis toward causal inference, applying Marcos López de Prado's 3-step framework — the same methodology used at ADIA Lab (Abu Dhabi's sovereign wealth fund). This lets us separate genuine causes from spurious correlations.

02 — Methodology

López de Prado's 3-Step Framework

Most finance papers stop at Step 1 (correlation). We go all the way to Step 3 (falsification).

STEP 1

Phenomenological

Observe statistical associations across 34 features and 4 pillars. ADF + KPSS stationarity tests. Correlation heatmap reveals unstable associations.

→

STEP 2

Theoretical

Propose causal structures using 4 complementary algorithms. Consensus rule: feature detected by ≥ 2/4 methods qualifies as causal candidate.

→

STEP 3

Falsification

Test causal claims via DoWhy: placebo treatment, random common cause, data subset refutation. Only robust features retained → sparse, validated DAG.

03 — Data Architecture

Four-Pillar Feature Organization

34 features from Bloomberg and CoinMetrics (2015–2025), organized into four thematic families.

On-Chain

9 features

CapMVRVHashRateMining Diff.Miner Rev.TxCntAdrActCntTX VolumeNVTSplyCur

Macro

12 features

DXYVIXMOVESPXCCMPRTYUS 3MGoldOilM2FARBAST2Y10

Sentiment

5+ features

Fear & GreedStablecoin MCapExchange Bal.Google "Bitcoin"Google "Coinbase"

Crypto Market

5 features

ETHBNBXRPLTCCFTC Futures

Preprocessing: log-returns for prices, first-differences for bounded ratios · forward-fill only (no look-ahead bias) · exclusion of BTC-derived features · weekly resampling for Google Trends compatibility

04 — Discovery Algorithms

Robustness via Multi-Method Consensus

Four independent algorithms with different assumptions, applied to the same weekly-resampled dataset.

PC Algorithm Constraint-based

Tests conditional independence via Fisher-Z on partial correlations. Removes edges that are conditionally independent. Contemporaneous structure only.

X ⊥⊥ Y | Z ⟹ remove edge

α=0.05 · 4 causes → BTC (ETH, BNB, CapMVRV, Miner Rev.)

NOTEARS Optimization

Continuous optimization: learns a weighted adjacency matrix W. Produces sparse DAG directly. Most aggressive filter — fewest edges retained.

min ‖X − XW‖² s.t. tr(e^W∘W) − d = 0

λ₁=0.1 · 2 edges (BTC→ETH, BTC→CapMVRV)

PCMCI Temporal

Partial correlations at each lag, conditioning on all other lagged variables. Separates contemporaneous from lagged effects — captures time structure.

ρ(X_t−τ, Y_t | Z) for τ ∈ {0, …, τ_max}

τ_max=4w · 26 significant links (most detections)

Granger Causality Predictive

F-test on nested VAR models — does X's past improve Y's forecast? Bidirectional testing. Sensitive to lag selection.

F-test: VAR(Y|Y_lag) vs VAR(Y|Y_lag,X_lag)

max_lag=8w · Many bidirectional links

Consensus rule: ≥ 2 out of 4 methods → 10 candidates from 34 tested.
Each then undergoes DoWhy falsification (placebo, random common cause, data subset).

05 — Validated Results

Final Causal DAG

10 features survived ≥ 2/4 discovery methods + DoWhy refutation tests. Zero Macro variables survived.

Method Consensus Matrix

Feature	Category	PC	NOTEARS	PCMCI	Granger	Count	DoWhy
CapMVRV	On-Chain	✓	✓	✓	—	3	✓
Miner Revenue	On-Chain	✓	—	✓	✓	3	✓
Mining Difficulty	On-Chain	—	✓	✓	—	2	✓
Est. TX Volume	On-Chain	—	—	✓	✓	2	✓
ETH	Crypto	✓	✓	✓	—	3	✓
BNB	Crypto	✓	—	✓	✓	3	✓
XRP	Crypto	—	—	✓	✓	2	✓
Stablecoin MCap	Sentiment	—	—	✓	✓	2	✓
Google "Coinbase"	Sentiment	—	—	✓	✓	2	✓
Google "Bitcoin"	Sentiment	—	✓	✓	—	2	✓

Eliminated: DXY, VIX, SPX — each detected by only 1/4 methods. Zero Macro variables survived.

Validated Causal Structure — 10 Drivers → BTC

Explore the Full Interactive Causal Graph (PC Algorithm)

Interactive Plotly graph — hover over nodes and edges to explore the complete causal structure discovered by the PC method.

06 — Multi-Horizon Analysis

No Universal Driver

Dominant factor families rotate across investment horizons — single-specification models are misleading.

Clustered MDA Importance by Family

Normalized positive Mean Decrease Accuracy (%) from Random Forest per pillar. Source: horizon_analysis.py

On-Chain

Macro

Sentiment

Crypto Mkt

T+1

49.2%

27.2%

23.0%

0.6%

T+7

15.7%

10.2%

64.6%

9.5%

T+30

26.2%

57.3%

1.8%

14.7%

On-Chain dominates at T+1 (49.2%) → Sentiment peaks at T+7 (64.6%) → Macro emerges at T+30 (57.3%)
No single pillar dominates universally. Driver rotation is the core finding.

07 — Predictive Validation

Causal vs. Full Feature Set

A Random Forest classifier (Bear / Neutral / Bull) demonstrates that causal features act as natural regularization.

Horizon	All Variables (138)	Best Causal (4–8)	Dummy Baseline
1d	33.6%	33.8%	33.0%
7d	33.1%	41.7%	30.3%
14d	34.7%	42.2%	28.8%
30d	31.9%	35.4%	22.7%
60d	18.1%	50.2%	12.9%

📉 Why 138 features fail

With ~400 weekly observations, the all-features model suffers from the curse of dimensionality: it overfits noise rather than learning signal. At 60d, it degrades to 18.1% — worse than the 12.9% dummy baseline.

📈 Why 4 features win

López de Prado's pipeline acts as domain-knowledge-driven feature selection. The 4-feature causal model avoids overfitting, reaching 50.2% balanced accuracy at 60d. At 7d: 66% Bear recall, 54% Bull recall.

08 — Regime Analysis

Macro Regimes as Exposure Filters

Although no Macro variable survived the causal pipeline, macro conditions still condition BTC returns — they're context, not cause.

BTC Forward Returns by Macro Regime (VIX × DXY median split)

0.1%

0.5%

Risk-on
Weak $

0.3%

0.5%

Risk-on
Strong $

0.4%

0.7%

Risk-off
Weak $

0.2%

0.7%

Risk-off
Strong $

14d returns 30d returns

30d returns are consistently higher than 14d across all regimes. Risk-off / weak USD shows the highest returns — contradicting a simple pro-cyclical narrative. This regime framework is designed as an intermediate step: as institutional investors enter the crypto market, new causal links to macro variables may emerge.

09 — Decision Framework

Operationalizing Causal Findings

A committee-ready decision process separating regime context from causal drivers.

1

Choose Horizon
Driver families rotate across horizons (Fig. 2). Select target investment horizon first.

▼

2

Apply Macro Regime Filter
VIX × DXY classification → context, not cause. Conditions the environment.

▼

3

Monitor Causal Drivers
On-chain + crypto market signals from the validated DAG. Transparent, explainable.

Output: Transparent, explainable exposure rules compatible with institutional governance.

10 — Conclusions

Key Findings

▸

No universal driver: Dominant factors rotate across horizons — single-specification models are misleading. On-Chain at T+1, Sentiment at T+7, Macro at T+30.

▸

Macro variables fail falsification: DXY, VIX, SPX each detected by only 1/4 methods. No Macro feature survives the full pipeline. This likely reflects the dominance of retail investors who don't trade BTC based on macro aggregates.

▸

On-chain + crypto market dominate: 7/10 validated drivers are blockchain-native (CapMVRV, mining metrics, altcoins). Bitcoin's price is driven by its own ecosystem.

▸

Altcoins are causal, not just correlated: ETH, BNB, XRP survive all 3+ methods and DoWhy refutation — their relationship with BTC passes causal falsification.

▸

Causal > full feature set: 4–8 causal variables outperform 138 features at all horizons beyond 1d. At 60d: 50.2% vs 18.1% balanced accuracy. Natural regularization through domain knowledge.

11 — Perspectives

Future Directions

🏛️ Institutional Adoption

The current causal structure reflects a retail-dominated market. As institutional players enter (ETFs, sovereign funds), new causal links to macro-financial variables — particularly liquidity proxies (Fed balance sheet, M2) — are likely to emerge, reshaping the DAG.

🔬 Non-Linear Methods

All 4 discovery algorithms assume linearity (Fisher-Z, linear SEM). Threshold effects and regime switches may require non-linear extensions such as kernel-based PCMCI.

⏱️ Temporal Resolution

Weekly resampling was necessary for Google Trends compatibility but loses intra-week dynamics. Daily-frequency causal discovery could reveal short-lived trading signals.

🧪 Out-of-Sample Testing

Validate the regime-aware framework on post-sample data and across other crypto-assets (ETH, SOL). Explore alternative regime definitions (growth × inflation à la Dalio's All Weather).