How it works

Quorvm is a closed-loop investment engine.

Six tiers turn raw market data into one auditable verdict per ticker. Every decision is hashed. Every gate vote is logged. Every persona’s contribution is attributable.

See it run live Honest limitations & disclaimer

System architecture · operator’s view

How 3,000 stocks become 1 executable verdict

Six tiers. 248 features. 17 personas. Every decision hashed. Every gate auditable.

Tier 01Data sources8 streams · price, fundamentals, sentiment, microstructure, regulatory

API

yfinance

Prices · fundamentals · technicals

API

Finnhub

News · earnings · realtime quotes

API

Alpha Vantage

Earnings · estimates

Reddit / X

Sentiment · momentum

API

NewsAPI

60+ outlets

API

IEX Cloud

Options chain · IV

CRON

Form 4 / 13F

Insider · institutional

INT

User watchlist

Tickers · prefs

Tier 02Feature engineering248 numerical features · materialized daily · indexed for the committee

p99 read 12ms · 4.2 GB/day write

18 FEAT

Momentum

12-1EARN DRIFTVOL-ADJ

24 FEAT

Sentiment

NLPBERTTOPICTONE

14 FEAT

Macro regime

YIELDVIXDXYCYCLE

32 FEAT

Microstructure

SPREADDEPTHIMB

40 FEAT

Alt data

DARKINSIDERFLOW

Tier 03ML models5 specialist models · each a discriminator the committee can interrogate

ML MODELAUC0.71

Transformer NLP

Earnings call tone

ML MODELPREC0.62

Anomaly detector

Insider · flow

ML MODELACC0.78

Regime classifier

5-state HMM

ML MODELIR1.84

Risk model

Multi-factor cov

ML MODELIC0.06

Alpha ensemble

GBM + LSTM stack

Tier 04Committee + adversarial debate17 personas · 3-round protocol · fail-loud on dissent

Mixtral-8x22B + Sonnet 4 + GPT-4o · 4–12s/case · ~$0.18/decision

QUANT

Quant ensemble

Alpha-Net v4 · GBM · LSTM

3 personasvote 0.20

MACRO

Macro

Dalio · Druckenmiller · Soros

3 personasvote 0.20

BEHAV

Behavioral

Lynch · Wood · contrarian

4 personasvote 0.20

VALUE

Value

Buffett · Munger · Graham · Burry

4 personasvote 0.20

TECH

Technical

Patterns · momentum · flow

3 personasvote 0.20

DEBATE

3-round adversarial protocol

1. Independent verdict (no peer info) · 2. Cross-examination, dissents flagged · 3. Final weighted vote, provenance hashed

Tier 05Sizing + hedge layerVerdict to dollars: Kelly capped, vol-targeted, hedge-overlaid

BARRA

Risk budget

Expected shortfall · vol · liquidity

COVID'08'22

KELLY

Position sizing

Fractional Kelly 0.5× · conviction-weighted · σ = 14% · min $500

HEDGE

Hedging layer

Earnings-week puts · sector ETF shorts · VIX call spread · tail-risk overlay

Tier 06Execution + auditBracket orders, decision log, post-trade attribution

GATES

12 pre-trade gates

Circuit breaker · earnings · liquidity · scenario · MC · red-team

EXEC

Alpaca bracket order

Auto stop-loss + take-profit at submission

TCA

Post-trade attribution

IS decomposition · slippage-adjusted IC

LOG

Immutable decision log

Every gate vote + force-bypass audited

3,000+

Universe

248

Features

Models

Personas

4–12s

Latency

$0.18

Cost / decision

Why each layer exists

In Arya’s words. No marketing.

Tier 01

Why 8 sources and not 1

Single-source quant systems collapse the moment that source ships a breaking change or rate-limits. Each Quorvm source is independently mockable, retryable, and covers a different signal type — losing any one degrades the system but does not stop it. Cross-source agreement (price + filings + flow) is itself a feature.

Rejected alternatives

✗Bloomberg Terminal API — $24K/year for one user. Defeats the cost moat.
✗Polygon-only — No fundamentals, no sentiment, no insider data.

Tier 02

Why 248 features and not 50

Signal diversity beats signal strength. A handful of "great" features collapses under regime change — momentum dies in 2008, sentiment dies in 2022. 248 features across 5 orthogonal categories means at least one category survives any regime, and the regime classifier (Tier 3) re-weights them on the fly.

Rejected alternatives

✗Ten "elite" features — High in-sample fit, brittle out of sample. Classic overfit trap.
✗Raw price only — Wastes everything fundamental and behavioral. Equivalent to buying SPY blind.

Tier 03

Why 5 specialist models and not 1 black box

Each model has a job. Each model can be evaluated independently. When the committee disagrees, you can ask exactly which model voted and why — versus a single "what does GPT think" prompt that gives you nothing to debug. Specialist models also age out gracefully: if regime classifier accuracy drops below 0.70, retire it without touching the others.

Rejected alternatives

✗One end-to-end transformer — Black box, expensive to retrain, no per-component IC.
✗Hand-coded rules only — Cannot adapt to new patterns. Stops working on the first regime change.

Tier 04

Why an adversarial committee, not one LLM

Single-LLM verdicts hallucinate confidently. The 3-round protocol forces every persona to produce an independent verdict before seeing peers, then explicitly cross-examines dissent, then hashes the provenance so you can audit which persona moved which way. Confident dissent is a red flag, not noise.

Rejected alternatives

✗Single Claude prompt — No dissent surfaces. Cannot tell if model is bullish or just agreeable.
✗Majority vote, no debate — Loses minority signal. Burry-style contrarians get drowned by the mode.

Tier 05

Why Kelly + hedge, not naive sizing

Naive equal-weight ignores conviction; full Kelly blows up on the first bad sequence. Half-Kelly captures most of the geometric growth without the ruin probability. The hedge layer is non-negotiable: every position carries an explicit tail-risk offset, sized against a 2008/2020/2022 stress scenario.

Rejected alternatives

✗Equal-weight 1/N — High-conviction picks get the same dollars as low-conviction picks. Wastes the committee.
✗Full Kelly — One bad streak triggers ruin. Kelly was always meant to be fractionalized in practice.

Tier 06

Why every trade has a hash trail

A signal that cannot be audited cannot be improved. Every entry passes 12 gates whose votes are persisted; every exit is reconciled against an arrival-price benchmark; every persona's contribution to the verdict is hashed. After 100 trades you can answer "which gate caught the most bad ideas?" — that's the closed loop.

Proof

Factor attribution + confidence calibration

The receipts. Per-factor IC. Per-persona calibration. The number that proves the AI works.

View the proof

Track record

Transparent · building since day one

Every recommendation logged. Every outcome scored. No cherry-picking, no survivorship.

See the record

What this is NOT

Not a guarantee. The committee can be wrong, frequently and confidently. The hash trail makes that wrongness measurable, not preventable.
Not high-frequency. 4–12s per decision means this is a swing/position engine, not a market-maker. Don’t arbitrage with it.
Not regime-proof. 5-state HMM degrades on novel regimes (think March 2020 week 1). The autoresearch loop is what closes that gap, slowly.
Not advice. This is a personal research tool. Read /disclaimer.

Analyzing stocks...