Complementing Behavioral Models in Capital Markets Execution & Settlement with LangGraph

In capital markets, collaborative filtering (CF) and behavioral models excel at pattern recognition: "Traders who executed NVDA blocks like this typically used Implementation Shortfall." But they fail catastrophically at causal reasoning: why that algo worked, whether current market microstructure supports it, and how clearing/settlement constraints modify the optimal choice. CF sees correlations; semantic retrieval understands mechanisms. This article demonstrates how to architect a hybrid recommender where semantic RAG provides the explanatory, constraint-aware reasoning layer that behavioral models lack—unified within a single LangGraph multi-agent system spanning both Order Execution and Clearing & Settlement modules.

The Real-Time Use Case: Cross-Module Execution-to-Settlement Advisor

Scenario: A trader needs to execute a €45M block of French equities while simultaneously ensuring T+2 settlement feasibility across Euroclear and domestic CSDs. Query: "Optimal execution strategy for BNP Paribas 600K shares considering current liquidity, upcoming ex-dividend date, and our failing trade exposure at Euroclear?"

Why Pure Behavioral Models Fail Here

• CF recommends algos based on historical fills but ignores that tomorrow is ex-dividend (settlement risk changes fundamentally).

• Behavioral models learn trader preferences but can't reason about CCP margin requirements or CSD cut-off times.

• Neither module communicates: execution choices create settlement obligations that may breach collateral limits.

Why Pure Semantic RAG Fails Here

• Retrieves relevant policies and market context but lacks personalized execution patterns from this specific trader/desk.

• Can't leverage years of fill data to calibrate which strategies actually minimize implementation shortfall for this name.

• Too slow and unstructured for real-time pre-trade decisions.

The Solution: A hybrid architecture where behavioral models generate candidate recommendations and semantic RAG validates, constrains, and explains them using real-time market state, regulatory corpus, and settlement infrastructure knowledge—all orchestrated by LangGraph with persistent cross-module state.

Part 1: The Hybrid Architecture Principle

Key Insight: Behavioral models answer "what has worked?" Semantic RAG answers "what should work now, given current constraints, and why?" The intersection is actionable, compliant, explained recommendations.

Part 2: Multi-Collection ChromaDB Architecture

Collection 1: Market Microstructure & Liquidity State (market_state)

Real-time updated embeddings capturing current trading conditions.

market_document = {
    "id": "mkt_bnp_20260812_0930",
    "page_content": "BNP Paribas Euronext Paris: Spread 3.2bps, TOB depth €180K, 
                     30d ADV €12M. Dark pool midpoint: €420K visible. 
                     Implied vol term structure backwardated. 
                     Ex-dividend date: 2026-08-13 (tomorrow). 
                     Record date settlement requires T+1 delivery.",
    "metadata": {
        "type": "market_state",
        "symbol": "BNP.PA",
        "venue": "XPAR",
        "spread_bps": 3.2,
        "tob_depth_eur": 180000,
        "adv_30d_eur": 12000000,
        "dark_midpoint_liq_eur": 420000,
        "ex_dividend_date": "2026-08-13",
        "record_date_settlement": "T+1",
        "iv_term_structure": "backwardated",
        "timestamp": "2026-08-12T09:30:00Z"
    }
}

Collection 2: Clearing & Settlement Infrastructure (settlement_infra)

Static + slowly-changing knowledge about CSDs, CCPs, cut-offs, and collateral rules.

settlement_document = {
    "id": "csd_euroclear_france_equity_t2",
    "page_content": "Euroclear France equity settlement: T+2 standard cycle. 
                     Cut-off for same-day matching: 16:00 CET. 
                     Failing trade penalty: 0.5bps/day after grace period. 
                     Collateral haircuts: French blue-chips 2%, mid-cap 5%. 
                     Current firm failing exposure: €2.1M (within €5M limit). 
                     Ex-dividend trades require cash collateral top-up by 14:00 CET D-1.",
    "metadata": {
        "type": "settlement_rule",
        "csd": "Euroclear_France",
        "asset_class": "equity",
        "settlement_cycle": "T+2",
        "matching_cutoff_cet": "16:00",
        "fail_penalty_bps_per_day": 0.5,
        "collateral_haircut_bluechip_pct": 2.0,
        "firm_fail_exposure_current_eur": 2100000,
        "firm_fail_limit_eur": 5000000,
        "ex_div_collateral_deadline_cet": "14:00",
        "effective_date": "2026-01-15"
    }
}

Collection 3: Regulatory & Best Execution Corpus (regulatory_corpus)

MiFID II RTS 27/28, SEC Rule 605, venue-specific obligations.

reg_document = {
    "id": "reg_mifid2_rts27_best_exec_large_cap",
    "page_content": "MiFID II RTS 27: For large-cap liquid equities, execution venues 
                     must be assessed quarterly on price, costs, speed, likelihood of execution. 
                     Pre-trade transparency waiver available for orders >€500K if reference price 
                     derived from regulated market. Post-trade deferral: 4 weeks for large transactions.",
    "metadata": {
        "type": "regulation",
        "framework": "MiFID_II",
        "instrument": "RTS_27",
        "asset_class": "equity_large_cap",
        "pre_trade_waiver_threshold_eur": 500000,
        "post_trade_deferral_weeks": 4,
        "jurisdiction": "EU"
    }
}

Part 3: Unified State Schema Spanning Execution & Settlement

from typing import Annotated, List, Dict, Optional, Literalfrom typing_extensions import TypedDict
from langgraph.graph.message import add_messages
from langchain_core.documents import Document

class ExecSettleRecState(TypedDict):
    messages: Annotated[List, add_messages]
    order_id: str
    trader_id: str
    
    # === BEHAVIORAL MODEL OUTPUT ===
    cf_candidate_algos: List[Dict]           # From collaborative filtering
    cf_confidence: float
    behavioral_signals: Dict[str, float]     # Trader prefs, historical fill quality
    
    # === SEMANTIC RETRIEVAL OUTPUT ===
    market_context: List[Document]
    settlement_constraints: List[Document]
    regulatory_context: List[Document]
    
    # === CROSS-MODULE VALIDATION STATE ===
    settlement_feasible: bool
    collateral_impact_eur: Optional[float]
    ex_div_risk_flag: bool
    regulatory_compliant: bool
    hybrid_score: float                      # Combined behavioral + semantic score
    validation_reasoning: str
    
    # Output
    recommendation: Dict
    explanation: str                         # Human-readable rationale
    agent_trace: List[str]

Part 4: Multi-Agent Implementation

Agent A: Behavioral Candidate Generator

Runs first. Produces candidates from CF/sequential models. Fast, pattern-based, no reasoning.

def behavioral_candidate_node(state: ExecSettleRecState):
    """Collaborative filtering + trader preference model. Sub-10ms."""
    # In production: call your existing ML serving endpoint
    # This is a placeholder representing your behavioral model infrastructure
    candidates = behavioral_model.predict(
        symbol=extract_symbol(state["messages"][-1].content),
        order_size=extract_size(state["messages"][-1].content),
        trader_id=state["trader_id"],
        market_regime=get_current_regime()
    )
    
    return {
        "cf_candidate_algos": candidates["algos"],      # e.g., [{"algo": "IS_AGG", "score": 0.82}, ...]
        "cf_confidence": candidates["confidence"],
        "behavioral_signals": candidates["signals"],
        "agent_trace": [f"behavioral: {len(candidates['algos'])} candidates, conf={candidates['confidence']:.3f}"]
    }

Agent B: Semantic Context Retriever

Retrieves market state, settlement constraints, and regulations relevant to THIS order. Runs in parallel with or immediately after behavioral generation.

from langchain_chroma import Chroma
from langchain_openai import OpenAIEmbeddings

market_store = Chroma(collection_name="market_state", embedding_function=OpenAIEmbeddings())
settlement_store = Chroma(collection_name="settlement_infra", embedding_function=OpenAIEmbeddings())
reg_store = Chroma(collection_name="regulatory_corpus", embedding_function=OpenAIEmbeddings())

def semantic_retriever_node(state: ExecSettleRecState):
    query = state["messages"][-1].content
    symbol = extract_symbol(query)
    
    # Parallel retrieval across three collections
    market_docs = market_store.similarity_search(query, k=3, filter={"symbol": symbol})
    
    # Settlement retrieval keyed on symbol's primary listing venue + asset class
    settle_docs = settlement_store.similarity_search(
        query, k=5,
        filter={"asset_class": "equity"}
    )
    
    reg_docs = reg_store.similarity_search(query, k=3, filter={"asset_class": "equity_large_cap"})
    
    return {
        "market_context": market_docs,
        "settlement_constraints": settle_docs,
        "regulatory_context": reg_docs,
        "agent_trace": state["agent_trace"] + [
            f"semantic: market={len(market_docs)}, settlement={len(settle_docs)}, reg={len(reg_docs)}"
        ]
    }

Agent C: Cross-Module Validation & Hybrid Scoring Agent

This is the core innovation. It takes behavioral candidates and validates/constrains/explains them using semantic context. The LLM reasons across modules.

from langchain_openai import ChatOpenAI
from langchain_core.prompts import ChatPromptTemplate

llm = ChatOpenAI(model="gpt-4o-mini", temperature=0)

hybrid_validation_prompt = ChatPromptTemplate.from_messages([
    ("system", """You are a Capital Markets Execution & Settlement Analyst.
    
    TASK: Validate behavioral model candidates against real-time semantic context.
    
    BEHAVIORAL CANDIDATES (what historically worked):
    {candidates}
    Confidence: {cf_confidence}
    
    MARKET STATE (current conditions):
    {market}
    
    SETTLEMENT CONSTRAINTS (clearing/CSD rules):
    {settlement}
    
    REGULATORY REQUIREMENTS:
    {regulatory}
    
    FOR EACH CANDIDATE ALGO, EVALUATE:
    1. SETTLEMENT FEASIBILITY: Does execution timing align with CSD cut-offs? 
       Will resulting settlement obligation breach collateral/fail limits?
    2. EX-DIVIDEND RISK: If ex-div is imminent, does algo complete before record date?
    3. REGULATORY COMPLIANCE: Pre-trade waiver eligibility? Venue assessment satisfied?
    4. MARKET FIT: Does current spread/depth/vol support this algo's assumptions?
    
    COMPUTE HYBRID SCORE = 0.4 × cf_score + 0.6 × semantic_validation_score
    (Semantic weighted higher because current constraints trump historical patterns)
    
    Return JSON: {{
        "validated_recommendations": [
            {{"algo": str, "hybrid_score": float, "settlement_feasible": bool,
              "collateral_impact_eur": float|null, "ex_div_risk": bool,
              "regulatory_compliant": bool, "rationale": str}}
        ],
        "overall_feasibility": bool,
        "explanation": str
    }}"""),
    ("human", "{query}")
])

def hybrid_validator_node(state: ExecSettleRecState):
    chain = hybrid_validation_prompt | llm.with_structured_output(dict)
    
    result = chain.invoke({
        "candidates": state["cf_candidate_algos"],
        "cf_confidence": state["cf_confidence"],
        "market": "\n---\n".join(d.page_content for d in state["market_context"]),
        "settlement": "\n---\n".join(d.page_content for d in state["settlement_constraints"]),
        "regulatory": "\n---\n".join(d.page_content for d in state["regulatory_context"]),
        "query": state["messages"][-1].content
    })
    
    top_rec = result["validated_recommendations"][0] if result["validated_recommendations"] else None
    
    return {
        "settlement_feasible": result["overall_feasibility"],
        "collateral_impact_eur": top_rec.get("collateral_impact_eur") if top_rec else None,
        "ex_div_risk_flag": top_rec.get("ex_div_risk", False) if top_rec else True,
        "regulatory_compliant": top_rec.get("regulatory_compliant", False) if top_rec else False,
        "hybrid_score": top_rec.get("hybrid_score", 0.0) if top_rec else 0.0,
        "validation_reasoning": top_rec.get("rationale", "") if top_rec else "No valid candidates",
        "recommendation": top_rec,
        "explanation": result["explanation"],
        "agent_trace": state["agent_trace"] + [
            f"hybrid_validator: feasible={result['overall_feasibility']}, "
            f"score={top_rec.get('hybrid_score', 0):.3f}" if top_rec else "hybrid_validator: no valid rec"
        ]
    }

Agent D: Synthesizer with Cross-Module Explanation

Generates trader-facing output that explicitly bridges execution and settlement reasoning.

synth_prompt = ChatPromptTemplate.from_messages([
    ("system", """You are an Institutional Trading Advisor.
    
    Present the validated recommendation with EXPLICIT cross-module reasoning.
    
    STRUCTURE YOUR RESPONSE:
    1. RECOMMENDATION: Algo + parameters
    2. EXECUTION RATIONALE: Why this algo fits current market conditions 
       (reference behavioral confidence AND market state)
    3. SETTLEMENT IMPACT: How this execution affects clearing/settlement 
       (collateral, cut-offs, fail risk, ex-div timing)
    4. RISK DISCLOSURES: Any remaining risks or conditional warnings
    5. ALTERNATIVES CONSIDERED: Why other behavioral candidates were rejected
    
    Be precise with numbers. Cite specific settlement rules and market metrics.
    
    Recommendation: {rec}
    Explanation: {explanation}
    Market: {market}
    Settlement: {settlement}"""),
    ("human", "{query}")
])

def synthesizer_node(state: ExecSettleRecState):
    chain = synth_prompt | llm
    response = chain.invoke({
        "rec": state["recommendation"],
        "explanation": state["explanation"],
        "market": "\n".join(d.page_content for d in state["market_context"][:2]),
        "settlement": "\n".join(d.page_content for d in state["settlement_constraints"][:3]),
        "query": state["messages"][-1].content
    })
    return {"messages": [("assistant", response.content)]}

Part 5: Compiling the Hybrid Graph

from langgraph.graph import StateGraph, START, END
from langgraph.checkpoint.postgres import PostgresSaver

workflow = StateGraph(ExecSettleRecState)

workflow.add_node("behavioral_candidates", behavioral_candidate_node)
workflow.add_node("semantic_retrieve", semantic_retriever_node)
workflow.add_node("hybrid_validate", hybrid_validator_node)
workflow.add_node("synthesize", synthesizer_node)

workflow.add_edge(START, "behavioral_candidates")
workflow.add_edge("behavioral_candidates", "semantic_retrieve")
workflow.add_edge("semantic_retrieve", "hybrid_validate")

def validation_router(state: ExecSettleRecState):
    if state["settlement_feasible"] and state["regulatory_compliant"]:
        return "synthesize"
    # Even failures get synthesized with safe fallback explanation
    return "synthesize"

workflow.add_conditional_edges("hybrid_validate", validation_router)
workflow.add_edge("synthesize", END)

checkpointer = PostgresSaver.from_conn_string("postgresql://...")
app = workflow.compile(checkpointer=checkpointer)

Part 6: How Semantic Retrieval Complements Behavioral Models

Capability	Behavioral Model Alone	Semantic RAG Alone	Hybrid (This Architecture)
Personalization	✅ Learns trader/desk patterns	❌ Generic context only	✅ Personalized + constrained
Current Market Awareness	❌ Lagging indicators only	✅ Real-time microstructure	✅ Real-time validated against history
Settlement Constraint Reasoning	❌ Invisible to exec models	✅ Explicit infrastructure knowledge	✅ Cross-module feasibility scoring
Regulatory Compliance	❌ Cannot reason about rules	✅ Corpus-grounded validation	✅ Validated per-candidate
Ex-Dividend / Corporate Action	❌ Not in fill history features	✅ Calendar-aware retrieval	✅ Timing-feasible recommendations
Explainability	❌ Black-box scores	✅ Natural language reasoning	✅ Bridged explanation across modules
Latency	✅ Sub-ms	❌ 100ms+ retrieval + LLM	⚠️ 200-500ms (acceptable for advisory)

The complementarity is structural: Behavioral models provide the candidate generation prior. Semantic RAG provides the constraint satisfaction posterior. The hybrid score is the Bayesian update: P(optimal | history, current_constraints) ∝ P(history) × P(constraints | candidate).

Production Considerations

1. Latency Budget Allocation: Behavioral model: <10ms. Semantic retrieval: <100ms (pre-warm collections, cache hot symbols). Hybrid validation LLM: <300ms (use structured output, small model). Total p99 target: <500ms for advisory use case. For HFT-latency execution, skip LLM validation and use deterministic rule engine.

2. Market State Freshness: market_state collection must be updated via streaming pipeline (WebSocket/kafka) with TTL-based expiry. Stale market embeddings invalidate all hybrid scoring. Implement freshness checks in retriever node.

3. Settlement Data Governance: settlement_infra is quasi-static but critically versioned. Include effective_date, superseded_by, and source_document_ref metadata. Auto-expire stale rules. Audit every change.

4. Behavioral Model Drift Detection: Track hybrid_score vs. cf_score divergence over time. Persistent divergence means either behavioral model is stale OR semantic context is miscalibrated. Alert quant desk when divergence exceeds threshold.

5. Cross-Module State Persistence: Use Postgres checkpointer to maintain conversation memory spanning execution and settlement queries. A trader asking about settlement implications 10 minutes after execution advice should have full context without re-querying.

6. Evaluation Beyond Acceptance Rate: Measure (a) settlement fail rate reduction vs. behavioral-only baseline, (b) collateral optimization savings, (c) ex-div timing violation rate, (d) trader trust score via explicit feedback. These validate that semantic complementarity creates real value.

Conclusion

In capital markets, semantic retrieval doesn't replace behavioral models—it completes them. Collaborative filtering knows what worked yesterday. Semantic RAG knows what constraints apply today. The hybrid architecture presented here makes this complementarity operational: behavioral candidates flow through semantic validation, emerging as recommendations that are simultaneously personalized, current-market-aware, settlement-feasible, and regulatorily compliant. For Order Execution, semantic retrieval adds the why and the whether-now. For Clearing & Settlement, it adds the execution-origin awareness that pure settlement systems lack. United in a single LangGraph state machine, these modules stop being silos and become a coherent decision system—one that understands that every execution choice is also a settlement obligation, and every settlement constraint should inform execution strategy. That cross-module reasoning is where alpha lives and where risk hides. Neither behavioral models nor semantic RAG alone can capture it. Together, they can.