Skip to content

comparative_analysis_quant_frameworks.md

Latest commit

 

History

History
196 lines (164 loc) · 10.9 KB

comparative_analysis_quant_frameworks.md

File metadata and controls

196 lines (164 loc) · 10.9 KB

Comparative Analysis — Quant Agentic Frameworks vs agent-utilities

Comparison of three open-source quant agentic frameworks against the agent-utilities finance domain (agent_utilities/domains/finance/*, CONCEPT:KG-2.6), with three concrete gaps closed:

  • CONCEPT:KG-2.26 — Trade-journal bias auditor + shadow account
  • CONCEPT:KG-2.27 — Agent calibration / reputation tracking
  • CONCEPT:KG-2.28 — Persona decision-heuristic enrichment

The thesis: each borrowed capability is made stronger by our OWL/KG / graph-native / self-evolution substrate — the borrowed feature stops being a one-shot report and becomes a reasoned-over, provenance-bearing, calibrated fact the rest of the system can cite.

1. Per-repo capability summary

Vibe-Trading — research / strategy-generation studio

A skill-rich research agent: 74 skills (ML-strategy, SMC, Elliott-wave, sector-rotation, perp-funding-basis, on-chain analysis, liquidation-heatmap, options/SABR vol-surface, …) and 29 swarm presets. Its standout, novel capability is the shadow account / trade-journal audit (agent/src/shadow_account/*, agent/src/tools/trade_journal_tool.py): it parses a broker export, FIFO-matches roundtrips, builds a trading profile (win rate, holding days, PnL ratio, max drawdown), and runs 4 behavioural-bias diagnostics (disposition effect, overtrading, momentum-chasing, anchoring), then distils profitable roundtrips into if-then shadow rules and backtests the counterfactual "shadow" portfolio against the user's real trades for delta-PnL attribution. The audit is a flat report — not persisted as a queryable fact.

AutoHedge — thin autonomous crypto hedge-fund pipeline

A deliberately thin autonomous loop (autohedge/): a fixed Director → Quant → Risk → Execution agent chain (plus sentiment), structured JSON output, risk-first sizing, on-chain (Solana/Jupiter) execution. Strength: clean risk-first pipeline ergonomics. Weakness: no memory of which agent was right before, no behavioural feedback, single-venue, no graph substrate. Its role taxonomy already maps onto our SwarmRole.

FinceptTerminal — Bloomberg-class terminal

A desktop terminal (fincept-qt/) wrapping 100+ data sources (yfinance, OECD/BEA/BoE macro, akshare, financial-datasets, …), 16 broker execution integrations, and the richest persona registry of the three: 37 persona agents across TraderInvestorsAgent, hedgeFundAgents (Renaissance-style archetypes), GeopoliticsAgents, and EconomicAgents. Critically, each persona config (TraderInvestorsAgent/configs/agent_definitions.json) carries an explicit, scored decision framework — Buffett: ROE≥15% / D/E<0.5 / owner-earnings yield; Graham-style value gates; named scoring_weights and required line_items. That structure is the seed for KG-2.28. Weakness: no KG, no calibration/reputation, frameworks live in prose + weights but aren't executable, queryable graph facts.

2. Gap table

Capability Vibe-Trading AutoHedge FinceptTerminal agent-utilities (before) agent-utilities (after)
Bull/bear + persona-voice debate w/ risk veto partial basic persona prose ✅ debate_engine + investor_debate ✅
Role-weighted swarm consensus preset swarms fixed chain — ✅ trading_swarm ✅
Forensic earnings screen (M/Z/F/Sloan) — — data only ✅ forensic_screener (engine) ✅
Portfolio opt / VaR / regime / alpha factors partial basic data ✅ KG-2.6 suite ✅
Trade-journal behavioural-bias audit ✅ (flat report) — — ❌ ✅ KG-2.26 (KG-persisted)
Shadow account (trader profile as signal) ✅ (report) — — ❌ ✅ KG-2.26
Agent calibration / reputation feedback — — — ❌ ✅ KG-2.27 (Brier → swarm weights)
Executable, queryable persona heuristics — — prose + weights partial (voice only) ✅ KG-2.28 (OWL + evaluator)
100+ data sources / many brokers some 1 venue ✅ partial (out of scope)
Graph-native provenance / OWL reasoning — — — ✅ epistemic-graph + OWL ✅

The three gaps closed here are the ones that are both absent from us and multiplied by our substrate — not the breadth gaps (data-source/broker count), which are integration surface rather than capability.

3. Hidden value-adds — where our substrate makes a borrowed capability stronger

The three frameworks treat these features as terminal outputs. Because we have an epistemic graph + OWL ontology + calibration-feedback loop, the same feature becomes an input to further reasoning:

  • Trade-journal audit → learning signal (KG-2.26). Vibe-Trading produces a PDF-shaped report. We persist the profile + each bias as :TraderProfile and :BehavioralBias nodes (EXHIBITED_BY), so a future Bull/Bear debate or the risk officer can cite them: "this account exhibits a HIGH disposition effect → weight the bear's stop-loss discipline up." The audit becomes a fact reasoned over, with provenance, not a one-shot artifact.
  • Calibration feedback (KG-2.27). Palantir AIP and Fincept have personas but no memory of who was right. We record each persona's directional calls vs outcomes, score them with the engine's brier_score kernel, and feed the calibration back into the weighted SwarmConsensus so historically-accurate voices outvote the rest. Each score is an :AgentCalibration node (CALIBRATION_OF the agent) — a queryable reputation, e.g. "which persona has the best Brier on tech shorts?"
  • Persona heuristics as OWL facts (KG-2.28). Fincept's frameworks live in prose + weights. We make them :DecisionHeuristic OWL individuals (HEURISTIC_OF a persona :Agent) with a deterministic evaluator, so the graph can answer "which personas' value criteria does ACME pass?" and a Buffett bull cites the exact passing/failing rule. The engine's forensic kernels feed Burry's short triggers directly — borrowed structure, grounded in our reasoned-over numbers.

In every case the differentiator is the same: reasoned-over facts + KG provenance + calibration feedback turn a static borrowed feature into a self-improving one.

4. Implemented gaps — API + wiring

a. Trade-journal bias auditor + shadow account — CONCEPT:KG-2.26

agent_utilities/domains/finance/trade_journal.py

from agent_utilities.domains.finance import TradeJournalAuditor, Roundtrip
auditor = TradeJournalAuditor()
profile = auditor.audit("acct_42", roundtrips)   # win rate, PnL ratio, max DD, 4 biases
profile.bias("disposition_effect").severity      # low | medium | high
auditor.persist(profile, backend)                # -> :TraderProfile + :BehavioralBias KG nodes
  • Real maths: FIFO-matched roundtrips in; win rate, avg holding period, PnL ratio (avg-win / avg-|loss|), cumulative-equity max drawdown; disposition (loser/winner hold ratio), overtrading (busy-vs-quiet-day PnL gap), momentum-chasing (buys >3% above own prior buy), anchoring (<5% entry-price CV).
  • Wiring → KG: to_batch() emits the standard ExtractionBatch (GraphNode / EnrichmentEdge) and persist() writes it through the same write_batch → GraphBackend path every enrichment source uses. None backend degrades to a no-op so the audit runs fully offline.

b. Agent calibration / reputation tracking — CONCEPT:KG-2.27

agent_utilities/domains/finance/calibration_tracker.py

from agent_utilities.domains.finance import (
    CalibrationTracker, apply_calibration_to_swarm,
)
t = CalibrationTracker()
t.record_call("quant_01", direction=+1, confidence=0.85, subject="ACME")
t.record_outcome("quant_01", realized_direction=+1, subject="ACME")
t.score("quant_01")                       # accuracy + Brier + calibration in [0,1]
apply_calibration_to_swarm(swarm, t)      # LIVE wire: mutates swarm.config.role_weights
  • Brier: engine client.finance.brier_score when reachable, vetted local fallback mean((f-o)^2) offline; calibration = clamp(1 - 2·brier, 0, 1).
  • Wiring → SwarmConsensus: calibrated_role_weights() scales each role's base weight by the average calibration of its agents (floored, never zeroed); apply_calibration_to_swarm() writes them back into the live swarm's config.role_weights, which the existing TradingSwarm.analyze weighted aggregation already consumes — so the next analyze() lets a high-calibration quant outvote a miscalibrated sentiment agent. A live-path test proves a swarm's decision flips HOLD → BUY after calibration.
  • Wiring → KG: persist() writes :AgentCalibration nodes (CALIBRATION_OF).

c. Persona decision-heuristic enrichment — CONCEPT:KG-2.28

agent_utilities/domains/finance/persona_heuristics.py

from agent_utilities.domains.finance import evaluate_persona, evaluate_all
ev = evaluate_persona("graham_investor", {"pe": 11, "pb": 1.1, "margin_of_safety": 0.4})
ev.verdict        # bullish | neutral | bearish | insufficient_data
ev.citation()     # names the exact passing/failing rules
  • Structured heuristics: PERSONA_HEURISTICS attaches typed Heuristic rules to Graham (P/E<15, P/B<1.5, MoS≥30%), Buffett (ROE/ROIC/owner-earnings/ D/E), Burry (forensic short triggers on Sloan accruals / Beneish-M / Altman-Z), Damodaran (DCF), Druckenmiller (regime), and a Lynch PEG lens. The Burry lens is inverted: satisfying the triggers ⇒ bearish. Missing metrics yield unknown (never a silent pass).
  • Wiring → debate/screen: DebateContext gained a metrics field; DebateEngine.persona_heuristic_evidence() / _heuristic_block() fold each bound persona's verdict into its bull/bear prompt by default when metrics are present (generic path untouched otherwise). A live-path test asserts the Graham verdict + KG-2.28 marker appear in the prompt block.
  • Wiring → KG/OWL: persona_heuristics_batch() / seed_persona_heuristics() emit :DecisionHeuristic nodes (HEURISTIC_OF); matching OWL classes (:DecisionHeuristic, :AgentCalibration, :TraderProfile, :BehavioralBias) and object properties (:heuristicOf, :calibrationOf, :exhibitedBy) were added to agent_utilities/knowledge_graph/ontology_quant.ttl.

5. Concept IDs

ID Capability Borrowed from Module
KG-2.26 Trade-journal bias auditor + shadow account Vibe-Trading domains/finance/trade_journal.py
KG-2.27 Agent calibration / reputation tracking (novel; gap vs Palantir/Fincept) domains/finance/calibration_tracker.py
KG-2.28 Persona decision-heuristic enrichment FinceptTerminal domains/finance/persona_heuristics.py

All three degrade cleanly offline (lazy engine + optional KG backend), export via domains/finance/__init__.py, and are covered by tests under tests/unit/finance/.