Architecture & Compliance

---

Transparent by Design. Trusted for Institutional Decisions.

MACRA.AI is not a black box. It is built from the ground up to meet the transparency, auditability, and explainability standards expected by financial institutions, asset managers, central banks, and regulators.

Where many machine learning systems obscure their reasoning, MACRA.AI makes every assumption, probability, and forecast path traceable and defensible:

• Causal inference is explicit: The model architecture is grounded in macroeconomic theory using structured VAR blocks, PCA signal compression, and Bayesian updating—not opaque latent factors.
   
• Probabilities are scenario-based and version-controlled: Every forecast reflects real-time data and a timestamped configuration of priors, inputs, and outcomes.
   
• Narrative signals are scored, not hallucinated: MACRA uses fine-tuned LLMs with output control to quantify narrative tone without introducing freeform language or uncontrolled bias.
   
• Audit and compliance-ready: Forecasts are logged with full input/output traceability, supporting internal model risk governance and external validation (e.g. SR 11-7, ECB TRIM, PRA SS1/23).
   

This architecture enables institutional users to anticipate macro turning points with confidence, while preserving the rigor, transparency, and control demanded by regulated environments.

Purpose: Build a real-time, high-integrity macro dataset.

Process:

• Ingests 25+ macroeconomic, policy, credit, labor, sentiment, and global indicators
   
• Aligns mixed-frequency data, applies vintage tracking, and performs volatility-aware smoothing
   
• Flags revisions, outliers, and regime shifts for accurate real-time inference
  

Output: A synchronized, structured macro signal set ready for model input
 

Purpose: Calculate recession probabilities using transparent, theory-driven modelling.

Process:

• Compresses inputs using PCA into interpretable macro components
   
• Applies a 5-block VAR system to capture causal linkages across global shocks, domestic dynamics, credit propagation, sentiment lags, and labour conditions
   
• Bayesian priors incorporate real-time policy stance, tone indicators, and structural assumptions
   
• Recalculates posterior probabilities weekly or after key events (CPI, NFP, FOMC, SLOOS)
   

Output: Real-time conditional forecasts for NBER-defined, technical, and perceived recessions
 

Purpose: Capture nonlinear risk acceleration and regime transitions.

Process:

• A Bi-LSTM model trained on 2008–09, 2020, and 2022–25 periods identifies complex temporal interactions
   
• Flags early signs of structural breaks and nonlinear feedback (e.g. QT + credit squeeze + weak labor)
   
• Controls for false positives during noise-heavy transitions (e.g. brief yield curve inversions)
   

Output: Enhanced precision in forecasting recession onset and depth
 

Purpose: Quantify forward-looking tone and integrate qualitative signals into forecast logic.

Process:

• Parses Fed speeches, Beige Book, CEO surveys, and sentiment indices
   
• Fine-tuned LLMs generate structured tone embeddings (e.g., inflation concern, credit tightening, labor optimism)
   
• Embeddings flow into Bayesian priors and sentiment-lag VAR block (VAR3)
   
• Scenario planner enables natural-language prompt injection into model pathways
   

Output: Forecasts that integrate real-time policy and sentiment shifts—before the hard data moves
 

Purpose: Deliver actionable, explainable insights through real-time interfaces.

Process:

• Forecasts updated continuously with macro event triggers and recalibrated weekly
   
• Outputs recession probabilities (NBER, GDP-based, perception-based) and macro landing zones (soft, stagnant, shallow, hard)
   
• Supports scenario simulation of 0–3 Fed cuts, credit stress spikes, oil shocks, or custom narrative inputs
   
• Available via API, dashboard, CLI, or export with full audit trails

Output: Live macro risk intelligence tailored for portfolio stress testing, policy forecasting, and client reporting