Quant Platform Documentation¶
Welcome to the technical documentation for Quant Platform — a historical-first, live-targeted quantitative research, simulation, and execution infrastructure designed around mathematical determinism, event-driven causality, and institutional accounting rigor.
Architecture at a Glance¶
The platform is constructed as a modular monolith where each layer has strictly bounded responsibilities and explicit semantic contracts:
Market Data Ingest (Bybit Trades / L2)
↓
Data Plane & Storage Roots (Natural Partitioning, Parquet, Manifest Hashes)
↓
PostgreSQL Catalog & Lineage Tracking (DDL, DataGateway Boundary)
↓
Representations & Feature Engine (Candles D03, Footprints D04, Features E04/E06)
↓
Validation & Walk-Forward Engine (Causal Floor, Purged & Embargoed Folds)
↓
Strategy Policies & Composition (G01 StrategySpec, G02 Signals, G04 Session/Cooldown, G03 Risk/Sizing)
↓
DecisionIntent (Learner-Agnostic, ADR-0009)
↓
Unified Execution Engine (H01 Order State Machine, H02 Fees/Slippage, H03 OCO/Conflict Resolver)
↓
Historical Replay Runtime (H05 Deterministic ReplayEngine, Trace Fingerprint)
↓
Double-Entry Portfolio Ledger (H04 Ledger Transactions, Realized/Unrealized PnL, Mark-to-Market Equity)
Documentation Sections¶
| Section | Focus & Contents | Key Documents |
|---|---|---|
| Product & Roadmap | Product vision, capability taxonomy, and staged execution roadmap | Product · Roadmap · Capability Map |
| Target Architecture | System topology, storage lifecycle, and market data ingest architecture | Target Architecture · Storage Lifecycle · Ingest Architecture |
| Core Contracts | Frozen semantic contracts (§1 through §33), DataGateway interface, and conformity gates | Core Contracts · DataGateway · CandleDefinition |
| Architecture Decision Records | Chronological immutable architectural decisions (ADR-0001 through ADR-0046) | ADR Index |
| Engineering & Operations | Definition of Done, repository synchronization, and deployment procedures | Definition of Done · Repository Sync |
| Integration & Proofs | End-to-end acceptance proofs and reproducible golden replay evidence | Wave 4 Golden Replay Proof |
Core Non-Negotiable Invariants¶
-
Bitwise Determinism & Monotonic Time: All simulations advance time exclusively via simulated event timestamps (
Instant). Replays run on fixture market data produce bitwise-identicalReplayResultartifacts, order logs, fill sequences, and SHA-256 trace fingerprints. -
Causal Availability Floor: The platform enforces point-in-time correctness: $\(t_{\text{available}} \le t_{\text{decision}}\)$ Any attempt to access data with a future availability timestamp fails closed with a fatal exception (ADR-0006, ADR-0031).
-
Double-Entry Accounting Conservation: All financial math is executed in exact
Decimalarithmetic. Positions are accounted in hedge mode, strictly conserving balance sheets: $\(\text{Equity}_t = \text{Cash}_t + \sum_{i} \text{PositionValue}_{i,t}\)$ -
DataGateway Seam Isolation: Upper layers (Strategy, Execution, Portfolio, Replay) must consume market data exclusively through
DataGateway.scan()andDataGateway.read(), never directly accessing storage files or raw database tables (ADR-0019).
Machine Learning & RL Integration¶
Quant Platform is engineered to provide institutional foundations for statistical learning and artificial intelligence in quantitative trading:
- Zero Lookahead Bias: Mathematically enforced causal floors (\(t_{\text{available}} \le t_{\text{decision}}\)) protect training features from accidental future leakage.
- Purged & Embargoed Cross-Validation: Native walk-forward cross-validation folds with combinatorial embargo intervals (ADR-0031) and Deflated Sharpe Ratio / PBO statistical evaluations (ADR-0037).
- Learner-Agnostic Decision Interface:
DecisionIntent(ADR-0009) allows arbitrary predictive models (Gradient Boosting, PyTorch Deep Learning, Reinforcement Learning agents) to emit trade signals without entangling execution state machines or accounting rules. - Sister Research Ecosystem: Recurrent neural networks, sequence predictors, and deep learning experiments sharing this algorithmic foundation are explored in our companion repository Rnn_V0_1.
Research Frontiers & Community RFCs¶
Beyond our frozen core architecture, we actively explore cutting-edge quantitative ideas with researchers and the open-source community:
- RFC #182: Live Execution Adapters — Evaluating CCXT Pro vs. Native Async WebSockets (Bybit / Interactive Brokers) within our deterministic state machine.
- RFC #183: Causal Reinforcement Learning & Reward Function Design — Formulating cost-penalized differential Sharpe rewards and risk budgets in
DecisionIntent. - RFC #184: Interactive Order Flow & Footprint Visualizer — Designing a terminal TUI vs. lightweight web UI for cluster delta footprints and replay stepping.
- RFC #185: System One Decision Models (TypeSafe Jev) — Evaluating fast typed inference models against exchange fee hurdles and adverse-selection gating.
- RFC #186: JEPA World Models for Order Flow & L3/MBO Data — Self-supervised representation learning (Fin-JEPA) on unlabelled high-frequency market events.
Have an idea or research proposal? Join the debate in Discussions ➔ Ideas!