import React from 'react'; import { BookOpen, X, Cpu, BarChart2, Compass, ShieldAlert, Activity } from 'lucide-react'; import 'katex/dist/katex.min.css'; import { BlockMath, InlineMath } from 'react-katex'; interface CryptoMathModalProps { isOpen: boolean; onClose: () => void; } export default function CryptoMathModal({ isOpen, onClose }: CryptoMathModalProps) { React.useEffect(() => { const handleKeyDown = (e: KeyboardEvent) => { if (e.key === 'Escape') { onClose(); } }; if (isOpen) { window.addEventListener('keydown', handleKeyDown); } return () => { window.removeEventListener('keydown', handleKeyDown); }; }, [isOpen, onClose]); if (!isOpen) return null; return (
{/* Modal Header */}

Crypto Bayesian Markov - Math & Logic Specification

Institutional Specification Manual

{/* Modal Body */}

4. Crypto Bayesian Markov Engine

Models momentum regimes and updates transition probabilities using on-chain alpha inputs.

{/* Section A: Markov Chain State Space */}

A. Markov Chain State Space

The asset return state space is mapped into 3 momentum regimes:

State 1 () Bearish Squeeze / Crackdown
State 2 () Consolidation / Mean Reversion
State 3 () Parabolic Bull Run
{/* Section B: Transition Matrix (P) */}

B. Transition Matrix (P)

Calculates transition probabilities over rolling 90-day return vectors:

where represents the frequency probability of moving from State to State .

{/* Section C: Bayesian Update Engine */}

C. Bayesian Update Engine

When external alpha inputs (e.g. Funding Rate anomalies, Whale inflows) occur, state probabilities are updated using Bayes' theorem:

Where:
- is the prior state probability from the Markov transition matrix.
- is the conditional likelihood of observing this whale spike / funding squeeze in State .

{/* Section D: Two-Stage Hybrid Machine Learning Pipeline */}

D. Two-Stage Hybrid Machine Learning Pipeline

To minimize false signals and adapt to changing momentum regimes, we run a two-stage hybrid ML pipeline. First, a client-side Random Forest Classifier maps non-linear combinations of four features to forecast trend probabilities:

Ensemble Feature Inputs:
  • Funding Rates (): Measures leverage imbalance in futures contracts.
  • Open Interest Volatility (): Captures the contract buildup speed.
  • Long/Short Retail Skew (): Identifies retail sentiment extremes.
  • Whale Accumulation Inflows (): Proxy for cold-wallet transfer velocities.

The Random Forest ensemble evaluates 10 non-linear decision trees to output the short-term and medium-term trend probabilities:

{/* Section E: Conjugate Beta-Binomial Error-Correction Engine */}

E. Conjugate Beta-Binomial Error-Correction Engine

Raw ML forecasts are susceptible to regime drift. We implement a self-correcting Beta-Binomial update. The continuous probability output from the Random Forest model is mapped into a discrete Binomial likelihood by defining the Trust-Weight Hyperparameter () as the Effective Sample Size (ESS):

1. Prior Distribution (Accuracy History):

where represents historical prediction successes and represents false alarms.

2. Likelihood Formulation (Binomial pseudo-observations):

where is the number of simulated successes and is the number of simulated failures.

3. Conjugate Posterior Update:

{/* Section F: Mathematical Proof of Posterior Mean Integration */}

F. Mathematical Proof of Posterior Mean Integration

To resolve a single operational point-estimate from our posterior distribution, we integrate out the continuous parameter to calculate the mathematical expectation of the posterior distribution:

Posterior Expectation Integral Proof:

Using the definition of the Beta function and the recurrence relation :

Expanded Workstation Implementation Formula:

This formulation ensures that if the prior model is highly accurate (large ), the raw ML signal is smoothed towards historical baseline expectations. If historical errors are high, the prior variance restricts overreaction to noisy signals.

{/* Section G: Ensemble Estimator Specifications */}

G. Ensemble Estimator Specifications

The Walk-Forward Ensemble Radar aggregates forecasting signals from 5 independent machine learning estimators optimized for distinct predictive roles across three temporal horizons:

1. Random Forest (RF)

Utilizes bootstrap aggregation (bagging) of uncorrelated decision trees to map non-linear feature interactions. It constructs a robust ensemble prediction:

Optimized for multi-regime boundary separation and filtering out high-volatility futures noise.

2. XGBoost / Gradient Boosting (GB)

Fits sequential decision trees to minimize the residual classification loss via gradient descent:

Highly responsive to short-term micro-trends, making it the primary signal anchor for the T+1 horizon.

3. Logistic Regression with ElasticNet (LR)

Serves as the linear baseline anchor, regularized with combined L1 (Lasso) and L2 (Ridge) penalties:

Prevents wild regime-extrapolation decay and ensures structural stability during major trend shifts.

4. Support Vector Machine (SVM)

Projects the feature space into a high-dimensional Hilbert space using a Radial Basis Function (RBF) kernel:

Isolates non-linear hyperplane separation boundaries, targeting multi-dimensional trend-reversal thresholds for the T+5 horizon.

5. Multi-Layer Perceptron (MLP)

A deep feedforward neural network mapping complex cross-correlations across hidden layers using backpropagation:

Extracts intricate temporal patterns and deep feature interactions, optimized for the medium-term T+10 forecasting horizon.

); }