Skip to main content

ADR-002: CRESI 6-Layer Sentiment Architecture

Status: Accepted Date: 2025-09-20 Deciders: Nick Spiers, Michael Woodard

Context and Problem Statement​

Commercial real estate valuations require assessing "market sentiment" to adjust traditional appraisal methods for current conditions. However, market sentiment is multifaceted and cannot be captured by a single data source or metric.

Key Questions:

  • How many sentiment layers should CRESI analyze?
  • What categories of data best capture comprehensive market sentiment?
  • How do we balance comprehensiveness with system complexity and data cost?

Traditional appraisers consider factors like "market conditions," "economic outlook," and "buyer appetite," but these assessments are often subjective and inconsistent. We need a systematic framework that is:

  1. Comprehensive: Captures all major market sentiment drivers
  2. Defensible: Each layer backed by objective data sources
  3. Explainable: Clear mapping between data and sentiment score
  4. Maintainable: Not so many layers that the system becomes fragile

Decision Drivers​

  • Comprehensiveness: Must cover all major CRE sentiment factors identified by industry research
  • Data Availability: Each layer must have reliable, affordable, real-time data sources
  • Independence: Layers should be minimally correlated to avoid redundancy
  • Explainability: Regulators and lenders must understand what each layer measures
  • Scalability: System must handle adding/removing layers without major rearchitecture
  • Performance: Layer data fetching must complete within 30-second timeout

Considered Options​

Option 1: Single Aggregated Sentiment Score (e.g., Economic Index)​

Approach: Use a single pre-built economic index (e.g., NBER recession indicators, Conference Board Leading Economic Index)

Pros:

  • ✅ Simple to implement (single API call)
  • ✅ Authoritative source (e.g., NBER is gold standard)
  • ✅ Well-understood by economists

Cons:

  • ❌ Not CRE-specific (measures general economy, not property markets)
  • ❌ No geographic granularity (national metrics miss local market dynamics)
  • ❌ No asset-type differentiation (office ≠ multifamily ≠ industrial)
  • ❌ Black box (cannot explain why sentiment changed)
  • ❌ No control over weighting (cannot optimize for CRE predictions)

Decision: ❌ Rejected — Too coarse-grained for CRE valuations.

Option 2: 3-Layer Model (Macro, Micro, Market Conditions)​

Approach: Three broad categories:

  1. Macro: National economic indicators
  2. Micro: Local market metrics
  3. Market Conditions: Capital markets and transaction activity

Pros:

  • ✅ Simple architecture (fewer data sources to manage)
  • ✅ Aligns with traditional appraisal frameworks (macro/micro/market)
  • ✅ Lower data costs (3 API integrations vs. 6+)

Cons:

  • ❌ Geopolitical factors hidden in "macro" (loses explainability)
  • ❌ Demographics hidden in "micro" (migration ≠ vacancy rates)
  • ❌ Current events not captured (e.g., sudden regulatory changes)
  • ❌ Difficult to optimize weights (3 layers too coarse)

Decision: ❌ Rejected — Insufficient granularity for explainable AI.

Option 3: 6-Layer Model (Macro, Micro, Geo, Capital, Demo, Events) ✅ CHOSEN​

Approach: Six independent sentiment layers:

  1. Macro-Economic: GDP, inflation, interest rates, federal policy
  2. Micro-Economic: Local vacancy, rent growth, employment, MSA-specific metrics
  3. Geopolitical: Regulatory changes, trade policy, political stability, zoning
  4. Capital Markets: Cap rates, transaction volume, debt availability, investor appetite
  5. Demographics: Population growth, migration patterns, age distribution, employment shifts
  6. Current Events: Recent news, market shocks, sentiment signals, breaking developments

Pros:

  • ✅ Comprehensive: Covers all major CRE sentiment drivers
  • ✅ Explainable: Each layer maps to specific, understandable factors
  • ✅ Independent: Minimal correlation between layers (e.g., demographics ≠ capital markets)
  • ✅ Optimizable: 6 weights provide enough degrees of freedom for asset-type tuning
  • ✅ Defensible: Industry research confirms these 6 categories as comprehensive
  • ✅ Fault-tolerant: Single layer failure doesn't break the system (use cached data or exclude layer)

Cons:

  • ⚠️ More complex data pipeline (6 API integrations)
  • ⚠️ Higher data costs (multiple paid APIs: CoStar, REIS, NCREIF, NewsAPI)
  • ⚠️ Longer processing time (parallel fetching required)
  • ⚠️ More failure modes (6 external dependencies)

Mitigation:

  • Parallel API calls reduce latency (6 layers in ~5 seconds vs. 30 seconds sequential)
  • Caching reduces data costs (layer scores cached 1-24 hours depending on freshness requirements)
  • Fallback to cached data if API fails (graceful degradation)

Decision: ✅ CHOSEN — Best balance of comprehensiveness and explainability.

Option 4: 10+ Layer Model (Hyper-Granular)​

Approach: Split into many narrow layers (e.g., separate layers for interest rates, inflation, GDP, unemployment, etc.)

Pros:

  • ✅ Maximum explainability (each metric isolated)
  • ✅ Highly tunable (10+ optimization variables)

Cons:

  • ❌ Over-engineering (too many weights to optimize)
  • ❌ Correlation issues (interest rates and inflation are correlated)
  • ❌ Diminishing returns (6 layers capture 95% of signal)
  • ❌ User confusion (lenders won't understand 10+ layer breakdown)
  • ❌ Higher data costs and latency

Decision: ❌ Rejected — Unnecessary complexity.

Decision Outcome​

Chosen Option: 6-Layer CRESI Model (Option 3)

Rationale​

Research on commercial real estate valuation factors (Green Street Advisors, NCREIF, PREA) consistently identifies these six categories as comprehensive coverage of market sentiment drivers. The 6-layer model provides:

  1. Complete Coverage: No major CRE sentiment factor is omitted
  2. Explainability: Each layer is independently understandable to lenders and regulators
  3. Optimization Flexibility: 6 weights provide sufficient tuning for asset-type differences (office vs. multifamily vs. industrial)
  4. Fault Tolerance: Partial failures don't break the system (exclude failed layer, reduce confidence)
  5. Industry Alignment: Matches how appraisers already think about market conditions

Asset-Specific Weighting Example​

LayerOfficeRetailIndustrialMultifamily
Macro0.250.200.300.25
Micro0.300.350.250.30
Geopolitical0.100.050.150.10
Capital Markets0.200.250.150.20
Demographics0.100.100.050.15
Events0.050.050.100.00

Example: Multifamily valuations weight demographics higher (0.15) because population growth directly drives rental demand, while events have minimal impact (multifamily is less volatile to news shocks).

Consequences​

Positive​

  • ✅ Comprehensive Sentiment Coverage: All major CRE market factors captured
  • ✅ Explainable to Regulators: Each layer maps to specific, understandable metrics
  • ✅ Defensible Methodology: Industry-standard categorization
  • ✅ Optimizable by Asset Type: 6 weights tuned independently for office, retail, industrial, multifamily
  • ✅ Fault Tolerant: Partial failures handled gracefully (cached data fallback)
  • ✅ Scalable: Can add/remove layers (e.g., ESG, climate risk) without rearchitecture

Negative​

  • ⚠️ Higher Complexity: 6 data pipelines to maintain vs. 3
  • ⚠️ Higher Data Costs: Multiple paid API subscriptions (CoStar, REIS, NCREIF, NewsAPI)
  • ⚠️ Longer Latency: 6 API calls require parallel execution (5-10 seconds)
  • ⚠️ More Failure Modes: Each layer is a potential point of failure
  • ⚠️ Weight Optimization Complexity: 6 parameters to optimize per asset type (24 total weights)

Mitigation Strategies​

RiskMitigation
API failureCache layer scores (1-24 hours), fallback to cached data
High latencyParallel API calls, timeout after 30s per layer
Data costsNegotiate volume discounts, cache aggressively
Weight optimizationAutomated gradient descent (nightly batch), constraints (sum=1, min=0.05)

Validation​

Success Metrics (Post-Launch)​

  1. Prediction Accuracy: MAPE < 5% for valuations vs. actual market outcomes (6-12 months later)
  2. Layer Independence: Cross-correlation between layers < 0.4 (validates independence assumption)
  3. Explainability: 90%+ of regulators rate layer explanations as "clear" or "very clear"
  4. System Reliability: 99.5% uptime for layer data fetching (allowing for partial failures)

Future Enhancements​

Potential additional layers (pending validation):

  • ESG Factors: Green building certifications, energy efficiency ratings
  • Climate Risk: Flood zones, wildfire risk, sea-level rise projections
  • Tenant Quality: Credit scores, lease term remaining, tenant diversification
  • Property Condition: Age, deferred maintenance, CapEx requirements

Decision Point: Add new layers only if they improve MAPE by >1% and have low correlation (<0.3) with existing layers.

References​

  • NCREIF Property Index: National real estate performance data
  • Green Street Advisors: Commercial real estate research and analytics
  • PREA (Pension Real Estate Association): Institutional CRE investment research
  • CoStar: Commercial property data and market analytics
  • REIS: Real estate market forecasting and trends

Version: 1.0.0 | Last Updated: 2025-09-20