Regime Classifier
Four-state Gaussian HMM (GMM fallback) — low-volatility, mean-reverting, trending, high-volatility — auto-labeled by realized vol, with confidence-weighted hand-off between states.
Gaussian HMM/GMMConfidence-weighted4-state output
Built for desks like these
The engine
Not a dashboard. A cognition substrate.
ORDR Mind composes four reasoning layers — regime, volatility, order-flow, signal — into a single instruction stream. Every output is typed, signed and anchored. Every parameter is reproducible. Every decision is auditable.
The engine SynexFund is building for its own desk, exposed as a substrate your desk can wire into its existing OMS, risk and execution stack. No demoware. No metaphors for AI. The thing itself.
Cognition layers4 fused
Tick → signal latency< 50 ms target
Audit chainSHA-256 verifiable
Tenant isolationRLS + JWT + KMS
ORDR MIND · COGNITION ENGINE — REPRESENTATIVE VIEW
The platform
Six primitives. One coherent view of the market.
ORDR Mind is not a model. It is a composition of six reasoning layers, fused into a single instruction stream — every output cryptographically signed and routed through your existing execution stack.
Volatility Forecaster
Regime-aware volatility forecasting — the live regime conditions the forward envelope. Realized vol cone with a regime-derived bias.
Regime-aware1H · 4H · 1DVol cone
Order-Flow Inference
Order-flow and dealer-positioning features — gamma exposure, put/call, book imbalance. Surfaces the pressure that moves price (live options data vendor-gated).
L2 + L3MicrostructureData-gated
Autonomous Signals
Composite signal layer fuses regime, volatility and flow into a single executable instruction with cryptographic provenance hash.
Pydantic v2SignedMulti-agent DAG
Tamper-Evident Audit
Every model inference, signal emission and parameter change is appended to a SHA-256 hash chain — verifiable on demand, tenant-scoped, reproducible. No record can be silently altered.
SHA-256 chainverify_chainRLS-scoped
Multi-Market Coverage
Equities, FX, crypto, futures, sports, commodities, options — one cognition substrate, every market, normalised inference.
Multi-assetCross-assetCross-venue
Capability matrix
Every module. Every claim. The proof behind it.
Ten production modules, mapped to the capability they ship, the benefit your desk gets, and the test that proves it. No marketing without a receipt — where live data needs a premium vendor, we say so.
On the modules marked data-gated
The pipelines, feature stores and methodology are real, built and unit-tested — but live data requires a premium market-data vendor (e.g. an options feed for GEX). The rails are wired; the data plug is yours to bring. Everything else listed here is verified end-to-end in the test suite.
Methodology
From tick to signed instruction in four deterministic steps.
Every signal that leaves ORDR Mind can be traced back through this exact pipeline. No black boxes. No swapped models without lineage. No re-emission without a new hash.
01
Ingest
Streaming tick ingestion across venues
Kafka stream router with back-pressure. Polars zero-copy feature pipeline. Feast feature store. Designed for single-digit-millisecond tick-to-feature.
02
Infer
Three reasoning layers, fused
Regime classifier emits state with confidence. Volatility forecaster updates the envelope. Flow model decomposes order-book pressure. MLflow tracks every run.
03
Emit
Signed, typed, routed
Composite signal layer produces typed instructions with directional bias, sizing, confidence interval, SHA-256 evidence hash. Routed natively to your OMS/EMS.
04
Anchor
Hash-chained · tamper-evident · reproducible
Every inference batch is appended to a SHA-256 tamper-evident hash chain with on-demand verification. Your compliance team can reproduce and verify any decision the engine has ever made.
signal.flow / previewILLUSTRATIVE STREAM
PIPELINE · PROVENANCE
From raw tick to signed, audited inference.
How it actually works. No demoware, no synthetic model quietly promoted into production. Data is ingested, cached, backfilled point-in-time, trained, and only promoted if it survives an out-of-sample walk-forward gate — then served behind typed contracts and hash-chained on every inference.
pipeline.flow / data → model → serving → auditWALK-FORWARD GATED
01
INGESTStreaming tick → feature, in flight.
Kafka stream router · Polars feature pipeline · Feast feature store02
FEATURE CACHEFeatures written to hypertables, idempotently.
TimescaleDB · regime_features · realized_vol · asyncpg ON CONFLICT upsert03
BACKFILLRolling-window snapshots stamped at each bar's real timestamp.
Point-in-time history · no look-ahead · learnable time-series04
TRAINFit the regime model on real backfilled history.
scikit-learn · hmmlearn · HMM / GMM05
WALK-FORWARD GATEExpanding-window out-of-sample validation. Promote only if it beats a single-Gaussian baseline by margin.
The integrity gate · no synthetic model ever promotedINTEGRITY GATE06
REGISTRYVersioned artifact saved + promoted; loaded into serving at startup.
Model registry · honest stub served until a real model passes the gate07
SERVEServe the promoted model — or a transparent deterministic stub.
/forecast/regime · typed Pydantic contracts08
AUDITEvery inference, signal and parameter change appended to a tamper-evident chain.
SHA-256 hash chain · RLS-scoped per tenantMETHODOLOGY · STATISTICS
The science underneath the signal.
ORDR Mind is a statistical engine before it is a product. Regimes are inferred with a Gaussian Hidden Markov Model, validated out-of-sample with walk-forward folds, and shipped only when they beat a naive baseline on held-out likelihood. The rigor below is enforced in code — not asserted in a deck.
01 · Regime classification
A Gaussian HMM over four market states
A Gaussian Hidden Markov Model — with a Gaussian-Mixture fallback when covariances degenerate — segments the tape into four regime states, auto-labeled by ascending realized volatility: Low-Volatility, Mean-Reverting, Trending and High-Volatility. The matrix at right is the learned state-transition kernel — strongly diagonal, because regimes persist far more often than they flip.
FEATURE VECTOR ·vol_ratiotrend_strengthreturn_skewhurst_exponent
Low-VolatilityLOW-VOL
Mean-RevertingMEAN-REV
TrendingTREND
High-VolatilityHIGH-VOL
regime.transition_matrixREPRESENTATIVE
validation.walk_forwardREPRESENTATIVE
02 · Out-of-sample validation
Expanding-window walk-forward — train precedes test
The cardinal rule for financial time series is that training data must strictly precede the test block — no look-ahead, ever. Each fold trains on all history up to a cut and is scored on the next, unseen block. The train window expands with each fold, so later folds learn from more of the past while always being judged on the future.
A single random shuffle would leak tomorrow into today and inflate every metric. Expanding-window folds make leakage structurally impossible: the blue TRAIN segment is, by construction, a subset of the past.
03 · The promotion gate
Ship only what beats the baseline out-of-sample
A trained model is promoted to production only if it beats a single-Gaussian baseline out-of-sample by a required relative margin — default +2% — measured on held-out per-sample log-likelihood. A model that finds no generalizable structure is never shipped. For supervised models the metric is stratified k-fold cross-validation accuracy, never an in-sample fit.
PASS · +7.6% VS BASELINE · GATE +2%
The bars show held-out log-likelihood per fold: the HMM candidate (blue) sits consistently above the naive baseline (muted). Mean improvement clears the gate, so this candidate is eligible for promotion.
gate.held_out_log_likelihoodREPRESENTATIVE
None of this is marketing copy. The promotion gate is a real walk-forward backtest check in the codebase — a candidate that fails to beat the baseline out-of-sample is rejected before it can ever emit a signal.
Coverage
Cross-asset, cross-venue, single substrate.
One cognition layer, every market. Inference, correlation and signals normalised across asset classes — a regime shift in FX informs how you read crypto; a flow event in options informs how you size the underlying.
Equities
US · EU · JP · 4 venues
FX
Majors · EM crosses
Crypto
Spot · perps · options
Futures
Index · commodity · rate
Sports
MLB · NFL · NBA · soccer
Commodities
Energy · metals · ags
Cross-Market
Lead-lag correlation engine
Bespoke
Custom venues · private feeds
Cross-market correlation engine
A regime change in one asset is a leading indicator for another. ORDR Mind quantifies the lead-lag.
Technology
Production-grade, region-pinned, composable.
Each cognition layer exposes a typed interface. Regime, vol and flow models can be swapped, retrained or shadow-tested without disturbing the signal output. Every composition is reproducible — bit-identical, in any region.
- ●pre-prodInference runtimePython 3.13 · Pydantic v2 · FastAPI
- ●pre-prodFeature pipelinePolars · Feast · Kafka · Redis
- ●pre-prodModel registryMLflow · Prefect · Marquez lineage
- ●pre-prodStreaming busKafka 3.6 · exactly-once · MirrorMaker
- ●pre-prodAudit chainSHA-256 hash chain · verify_chain
- ●pre-prodMulti-tenant storeTimescaleDB RLS · ClickHouse · S3
- ●pre-prodIdentitySAML 2.0 · OIDC · JWT · RBAC
- ●pre-prodObservabilityOpenTelemetry · Prometheus · Grafana
1,608UNIT TESTS · 85% COV
<8msFEATURE PIPELINE · TARGET
<50msSIGNAL EMISSION · TARGET
ORDR MIND — COGNITION ARCHITECTURE v0.3.0
PROOF · SECURITY · COMPLIANCE
Built to be audited.
Institutions vet what they license. So these are not marketing estimates — they are measured from the engine's own test suite, and the security and financial rails below are implemented and tested today.
Engineering proof / measured
Figures are read directly from the codebase's own pytest run and coverage gate — not marketing estimates.
Security & compliance / control matrix
Honest scope: the security and financial rails described here are implemented and tested today. The live quant models — regime, volatility and order-flow inference — are gated on premium market-data vendor agreements and ship as those feeds come online. We label this plainly so nothing here is overstated.
The suite
Six modules. One licence. One contract surface.
Use the whole engine or pick the primitives your desk needs. Every module is exposed via the same typed contracts, the same RBAC layer, the same audit chain.