PINN Architecture & Flow Diagrams

Physics-Informed Neural Network System - Agencio Predict

1. System Overview

The PINN (Physics-Informed Neural Network) predictor is a trainable neural network that embeds trading physics constraints directly into the learning process. Unlike traditional ML approaches that apply guardrails post-prediction, PINN learns constrained dynamics from the start.

Key Advantages

2. Academic Foundations

The PINN implementation in Agencio Predict is grounded in peer-reviewed research from both machine learning and quantitative finance domains.

Key Academic References

3. Six PINN Phases

4. Neural Network Architecture

Model Parameters by Phase

5. Model Creation Process

6. Training Pipeline

7. Physics Loss Functions

Loss Weights by Phase

8. Feature Engineering

9. Prediction Flow

10. Correlation PINN (Phase 4)

Cholesky Parameterization

Correlation matrices must be Positive Semi-Definite (PSD). The Cholesky parameterization guarantees this by construction:

// Correlation matrix R = LL' where L is lower triangular
// L parameters: diagonal (softplus for positivity) + off-diagonal (unconstrained)

Cholesky L:          Correlation R = LL':
| l11  0    0  |     | 1       r12     r13     |
| l21  l22  0  |  →  | r12     1       r23     |
| l31  l32  l33|     | r13     r23     1       |

// Diagonal of R = sum of squares of L rows = 1 (normalized)
// Off-diagonal = dot product of L rows / norms

DCC (Dynamic Conditional Correlation) Physics

// Engle (2002) DCC model embedded as physics loss:
Q_t = (1 - a - b) × Q̄ + a × ε_{t-1} × ε'_{t-1} + b × Q_{t-1}

Where:
- Q̄ = unconditional correlation matrix (long-term average)
- ε_t = standardized residuals (returns / volatility)
- a = shock reaction parameter (how fast correlations react)
- b = persistence parameter (how long shocks last)
- Constraint: a + b < 1 for stationarity

Phase 4 Files

Phase 4 Test Coverage (94 tests)

11. Market Regime PINN (Phase 5)

Five Market Regimes

Crisis Response Actions

12. Cointegration PINN (Phase 6)

Cointegration PINN models the relationship between pairs of assets that move together in the long run (cointegrated pairs). It uses Ornstein-Uhlenbeck (OU) dynamics as physics constraints to learn mean-reverting spread behavior for pair trading strategies.

Ornstein-Uhlenbeck Physics Constraints

The OU process models mean-reverting dynamics: dS = θ(μ - S)dt + σdW

Pair Trading Signals

12 DSL Primitives (Phase 6)

Example: Pair Trading Strategy DSL

// GLD-SLV pair trading strategy using Cointegration PINN
{
  "universe": ["GLD", "SLV"],
  "entry": "and(
    pinn_is_cointegrated('GLD', 'SLV', 0.7),
    pinn_spread_in_entry_zone('GLD', 'SLV', 2),
    gte(pinn_pair_entry_signal('GLD', 'SLV'), 0.6)
  )",
  "exit": "or(
    pinn_spread_in_exit_zone('GLD', 'SLV', 0.5),
    gte(pinn_pair_exit_signal('GLD', 'SLV'), 0.7)
  )",
  "direction": "pinn_pair_direction('GLD', 'SLV')",
  "hedgeRatio": "pinn_hedge_ratio('GLD', 'SLV')"
}

Cointegration Model Files

Test Coverage

13. Algorithmic Trading Integration

Integration Points

Example: Using PINN in DSL Strategy

// Example DSL strategy using PINN primitives
{
  "entry": {
    "rules": [
      // Use Phase 1 Price PINN for direction
      { "fn": "pinn_is_bullish", "args": ["$symbol", 0.6] },
      // Use Phase 2 Volatility PINN for regime filter
      { "fn": "not", "args": [
        { "fn": "eq", "args": [
          { "fn": "pinn_vol_regime", "args": ["$symbol"] },
          "high_vol"
        ]}
      ]},
      // Use Phase 5 Regime PINN for crisis avoidance
      { "fn": "lt", "args": [
        { "fn": "pinn_crisis_warning" },
        0.3
      ]}
    ]
  },
  "sizing": {
    // Use Phase 3 Sizing PINN for optimal position
    "fn": "pinn_optimal_size",
    "args": ["$symbol", 30]  // 30 bps expected edge
  },
  "risk_controls": {
    "regimeSizingEnabled": true,  // Apply Phase 5 regime multiplier
    "maxPositionUsd": {
      "fn": "mul",
      "args": [
        10000,
        { "fn": "pinn_regime_risk_multiplier" }
      ]
    }
  }
}

Executor Crisis Response Flow

// From executor.ts - Crisis check using Phase 5 Regime PINN
const regimePrediction = macroContext.pinnData?.regimePrediction ?? null;
crisisCheck = await checkCrisisResponse(state.algorithmId, regimePrediction);

if (crisisCheck.shouldPauseEntries) {
  crisisPauseEntries = true;
  // Block new position entries
}

if (crisisCheck.shouldFlattenAll) {
  // Emergency: close all positions
  await flattenAllPositions(state, macroContext, 'crisis_flatten');
}

// Regime sizing multiplier applied to all new positions
if (riskControls?.regimeSizingEnabled !== false && macroContext.pinnData?.regimePrediction) {
  const regimeMultiplier = dslPinnRegimeRiskMultiplierSync(macroContext.pinnData);
  // Crisis: 0.4x, Bear: 0.6x, Recovery: 0.7x, Sideways: 0.9x, Bull: 1.1x
  positionSizeUsd *= regimeMultiplier;
}

13. DSL Primitives (50+)

Price PINN Primitives (Phase 1)

Volatility PINN Primitives (Phase 2)

Sizing PINN Primitives (Phase 3)

Correlation PINN Primitives (Phase 4)

Market Regime PINN Primitives (Phase 5)

14. Database Schema

15. API Endpoints

Admin Endpoints

Regime Endpoints

16. File Reference

Core Module

Phase-Specific Modules

Migrations

Validation Thresholds (Graduation Gates)

PINN Architecture Documentation - Agencio Predict
Generated: 2026-05-19 | All 6 Phases Complete | 150+ Unit Tests
Updated with Phase 4 Correlation PINN and Algorithmic Trading Integration