The Art Market Graph

1. The Art Market Data Problem

The art market is opaque by design. Unlike public financial markets with centralized reporting (Bloomberg Terminal, real-time quotes, mandatory disclosure), the art market operates on controlled information asymmetry: private gallery sales, paywalled auction databases, selective data release.

The same artist appears in dozens of name variants across sources. Building canonical records that deduplicate artist identities across data sources enables unified intelligence—but requires sophisticated resolution logic to balance speed, accuracy, and data quality across professional sources (news, auctions) vs. user input (autocomplete).

The Core Challenge: Detecting emerging artist momentum requires unified intelligence across auction data, news signals, institutional validation, and collector behavior—before the broader market recognizes the pattern. This is only possible with canonical artist identities that link disparate data sources with high accuracy and sub-second latency.

2. Multi-Source Data Ingestion

Egon operates a multi-source ingestion layer: news intelligence pipeline (daily RSS extraction from curated art publications), auction data integration (hybrid scraping with fallback methods), institutional validation APIs (museum collections with two-stage false positive filtering), and authoritative identity services (Getty ULAN for biographical data and name variants).

Tier-Aware Signal Weighting

Different signals carry different predictive weight depending on artist tier. This tier-aware relevance system is the key innovation enabling early momentum detection:

Weight Derivation: Multipliers derived from backtesting 2,000+ artist trajectories over 5-year periods. Discovery tier weights optimized for <18-month predictive horizon (tier elevation events), Blue-chip weights optimized for market timing (buy/hold/sell signals). Weights recalibrated quarterly based on prediction accuracy.

Freshness Weighting: Signal age affects relevance:

• Breaking (0-7 days): Highest multiplier (recent events weighted more heavily)
• Current (8-30 days): Baseline multiplier
• Evergreen (31+ days): Reduced multiplier with progressive decay over time

Signal Extraction: Each extracted signal undergoes Claude NLP analysis to determine base strength (1-10 scale), tier relevance classification, and freshness scoring. Tier-aware multipliers and temporal decay are then applied via a proprietary weighting algorithm to generate final actionable scores.

Data Universe Architecture

3. Canonical Artist Resolution System

The heart of Egon's data universe: a 7-tier matching pipeline achieving 99.98% link rate (4,980 AuctionArtist records linked, 1 orphan; 100% MarketSignal records linked).

Source-Aware Resolution Routing: The Latency/Accuracy Tradeoff

Authoritative verification provides high confidence but incurs latency. The key innovation: route authoritative lookups based on source type to optimize for speed (user autocomplete) or accuracy (professional data ingestion).

Data Freshness System (Automatic TTL Management)

Egon's sophisticated multi-tier TTL system automatically maintains data freshness across 10+ data sources without manual intervention:

Why Automatic TTL Enables Custom Deep Research:

• Autonomous data invalidation: System determines when data is stale without manual intervention. News signals refresh daily, museum data every 90 days, analysis cache at user-tier-specific intervals.
• On-demand LLM research: When cache expires, Claude performs extensive multi-query OSINT research + database enrichment—research depth impossible with static pre-cached data.
• Cost-performance balance: 85-98% cache hit rates reduce API costs (Getty ULAN, museum APIs, image collection) while preserving freshness where market timing matters (news, auction context).

Trigram Threshold Tuning (Optimized Balance)

The trigram similarity threshold balances precision (avoiding false matches) vs. recall (catching valid typos):

• Production threshold: High precision (>99%), strong recall (>94%) — current optimized setting
• Lower threshold: Reduced precision (~95%), higher recall (~98%) — increased false positives like "Pablo Picasso" matching "Pablo Neruda"
• Higher threshold: Extreme precision (>99.5%), reduced recall (~89%) — misses valid typos like "Pablo Piccaso"

Why Search Canonical Names Only: Searching aliases creates false positives (e.g., "Pablo" matches both "Pablo Picasso" canonical name and "P. Picasso" alias → ambiguous). Canonical-only search ensures 1:1 matching.

Confidence Scoring System

Each resolution tier assigns confidence scores used by downstream logic (investment grading, user recommendations):

Investment Grading Impact: Confidence < 0.7 downgrades investment grade by 1 letter (A+ → A). Confidence < 0.5 blocks investment grading entirely (returns "Insufficient Data"). See Investment Science whitepaper for complete grading methodology.

Performance Metrics (Production)

4. Cross-Dataset Intelligence

Canonical artist resolution enables sophisticated cross-dataset queries impossible with fragmented data. The key capability: single canonical ID links 10+ database tables for unified intelligence.

Universal Canonical Linking

Multi-Source Signal Aggregation

Confluence scoring requires data from 5 independent sources—impossible without canonical linking. For complete algorithm specification with weight derivation and contrasting examples, see Investment Science whitepaper Section 3.

# Conceptual illustration of multi-source confluence scoring
def calculate_confluence_score(canonical_id):
    # Single canonical ID enables cross-dataset queries
    signals = gather_signals_from_all_sources(canonical_id)

    # Apply proprietary tier-aware weighting
    weighted_signals = apply_tier_aware_weights(signals, artist_tier)

    # Multi-source confirmation bonus
    base_score = aggregate_weighted_signals(weighted_signals)
    confirmation_bonus = calculate_multi_source_bonus(len(signals))

    return normalize_to_scale(base_score * confirmation_bonus)  # 1-10 scale

Without canonical resolution: Would require querying each table with multiple name variants ("Pablo Picasso", "PICASSO, Pablo", "P. Picasso"), manual deduplication, potential double-counting, 10-100x slower execution.

Real-Time Deduplication

Canonical linking enables automatic deduplication across user activities. For example, a collector tracking "Pablo Picasso", "PICASSO, Pablo", and "P. Picasso" in their watchlist would see 3 separate entries without canonical resolution—but with canonical IDs, these automatically merge into a single artist record with unified engagement scoring.

5. Multi-Dimensional Vector Search

Egon creates 4 independent high-dimensional embeddings per canonical artist. The innovation: orthogonal dimensions enable weighted combination queries impossible on public web search.

4-Dimensional Embedding Architecture

Orthogonality Validation: Statistical correlation measured across thousands of artist pairs. Low correlation between dimensions confirms they capture independent artist attributes. Example: Aesthetic and market dimensions are largely independent—a lower-priced artist can have Blue-chip aesthetic quality (the basis for Discovery tier alpha opportunities).

Weighted Combination Queries

# Example 1: "Find artists aesthetically like Basquiat but cheaper"
search_results = vector_search_service.search_multi_dimensional(
    canonical_artist_id=basquiat_id,
    weights={
        'aesthetic': 'primary',    # Prioritize aesthetic similarity
        'market': 'secondary'      # Secondary factor: market positioning (cheaper = lower tier)
    },
    limit=10
)
# Returns: Artists with similar style (graffiti, neo-expressionism)
# but avg_hammer_price < $50K (vs Basquiat $500K+)

# Example 2: "Find Discovery tier artists with Blue-chip aesthetic"
search_results = vector_search_service.search_multi_dimensional(
    canonical_artist_id=blue_chip_artist_id,
    weights={
        'aesthetic': 'primary',      # Prioritize aesthetic match
        'investment': 'tertiary'      # Lower weight for investment profile (Discovery tier filter)
    },
    filters={'tier': 'Discovery'},
    limit=10
)
# Returns: Emerging artists ($5K-$25K) with established aesthetic appeal
# (undervalued opportunities before market recognition)

# Example 3: "Find artists with institutional validation but undervalued"
search_results = vector_search_service.search_multi_dimensional(
    canonical_artist_id=reference_artist_id,
    weights={
        'institutional': 'primary',    # Prioritize institutional similarity
        'market': 'secondary'          # Secondary factor: market position (undervalued filter)
    },
    filters={'avg_hammer_price__lt': 100000},
    limit=10
)
# Returns: Museum-validated artists (MoMA, Met, Tate) under $100K
# (institutional quality at accessible prices)

Why This Is Impossible on Public Web: Google/Bing cannot answer "aesthetically like X but cheaper" because they lack separate aesthetic and market embeddings. Web search collapses all dimensions into a single relevance score. Egon's 4-dimensional architecture enables precise control over which attributes matter for a given query.

Performance: Sub-200ms for weighted multi-dimensional search across 5,000+ canonical artists using optimized vector indexes.

6. Technical Architecture

Egon's data universe creates differentiated capabilities through three core technical innovations:

1. Source-Aware Resolution Routing

• Authority skip for autocomplete: Fast resolution for user input vs. thorough verification for professional sources
• Tiered caching strategy: Multiple cache tiers with TTLs optimized per data type for high cache hit rates
• Confidence scoring: Variable confidence ranges affect downstream investment grading (low confidence = grade penalty)

2. Tier-Aware Signal Weighting

• Discovery tier amplification: Gallery moves (highest weight), museum acquisitions (very high), retrospectives (high)
• Blue-chip tier dampening: Auction records (high weight), market shifts (moderate), milestones (minimal)
• Backtesting validation: Weights derived from 2,000+ artist trajectories, recalibrated quarterly

3. 4-Dimensional Orthogonal Embeddings

• Independent dimensions: Low correlation between dimensions ensures weighted queries control specific attributes
• Weighted combination queries: "Aesthetically like X but cheaper" impossible without orthogonal embeddings
• Sub-200ms vector search: Optimized vector indexes for fast similarity queries

Network Effects

More data sources → higher canonical link rate → better cross-dataset intelligence → more valuable insights → more users → more user interaction data → improved tier-aware signal weights

Key Differentiators

• ✅ 99.98% canonical link rate (4,980/1 auction records)
• ✅ Source-aware resolution routing (Getty skip for autocomplete, mandatory for professional sources)
• ✅ Tier-aware signal weighting (Discovery tier signals weighted significantly higher than Blue-chip expected performance)
• ✅ 4-dimensional orthogonal embeddings (r < 0.2 correlation enables weighted queries)
• ✅ Multi-dimensional semantic search ("aesthetically like X but cheaper" impossible on public web)
• ✅ Cross-dataset intelligence (5-signal confluence scoring across 10+ tables)
• ✅ Real-time deduplication (across all name variants via single canonical ID)