Deterministic Research.
Causal Trading Architecture.
Provable Alpha.
Most backtests lie because of subtle lookahead leakage, floating-point rounding drift, and naive fill assumptions.Quant.Platform (engineered by Neo-Nix) builds institutional-grade quantitative systems where every hypothesis is verified against monotonic causality, double-entry ledgers, and purged out-of-sample lockboxes.
A Causal Multi-Tier Pipeline
From raw uncompressed exchange feeds to deterministically reconciled double-entry equity. Each layer operates behind frozen contract interfaces, preventing semantic bleed.
DataGateway
Catalog-backed reader over partitioned Parquet storage. Monotonic timestamp cursor enforcing publication-delay floors.
Feature Registry
Strict point-in-time materialization. Eliminates forward-looking rolling calculations and normalizes continuous market states.
Policy Layer
Rule-based alphas, Supervised ML (Wave 5), and Causal RL policies (RFC #183). Generates typed intent without order book coupling.
Order Engine
Deterministic matching simulator. Applies realistic spreads, order book queues, slippage models, and exchange maker/taker fee tiers.
Double-Entry Core
Exact fixed-scale Decimal accounting. Immutable transaction journals preserving capital conservation laws down to fractional ticks.
The Fundamental Design Constraint
Alpha models cannot execute trades directly; Execution engines cannot modify alpha parameters; Ledgers cannot fabricate capital. All state mutations follow deterministic finite state machine (FSM) rules verified in CI.
The 5 Invariants of Causal Research
Over 90% of quantitative strategies fail when deployed live because academic backtests rely on unrealistic assumptions. In our lab, every research experiment must satisfy five non-negotiable invariants before capital consideration.
Strict Temporal Causality
No strategy can observe a price, candle close, or indicator until its broadcast timestamp has elapsed in real time. We strictly reject synthetic or forward-looking price arrays.
Purged & Embargoed Cross-Validation
Financial time-series possess long-memory serial autocorrelation. Standard k-fold cross-validation leaks information. We enforce temporal purge windows and post-trade embargoes between splits.
Exact Double-Entry Accounting
`float` arithmetic silently introduces compounding rounding errors over millions of simulated trades. Every fill, balance debit, and fee deduction is recorded using fixed-scale `Decimal` double-entry ledgering.
Frictional Reality & Fee Hurdle
Mid-price fills are an illusion. Orders are matched against real simulated book depth, factoring in maker rebates, taker penalties, queue priorities, and market impact slippage curves.
Out-of-Sample Lockbox
Tuning parameters until a curve looks pretty guarantees overfitting. Final model performance is verified once and once only on a cryptographically sealed lockbox partition of unseen market history.
Reproducible Provenance
Every backtest result records the exact commit hash, raw dataset SHA-256 fingerprint, seed, and environment dependencies for total determinism.
Research Tear Sheets & Case Studies
Quantitative hypotheses tested under audited institutional constraints. All metrics account for exchange fees, realistic slippage, and temporal lockbox verification.
Study 01: Causal Supervised ML on BTC
Multi-horizon volatility and momentum projections using strictly causal point-in-time features. Evaluated via purged walk-forward cross-validation with an untouched out-of-sample lockbox.
Study 02: Multi-Regime Volatility Breakout
Time-series momentum with regime-conditioned volatility filters and dynamic ATR position sizing. Formulated as a pure `StrategySpec` producing `DecisionIntent` actions.
Study 03: Order Flow Imbalance & VPIN
Measuring volume-synchronized probability of toxicity (VPIN) and order book queue asymmetry across L2/L3 feeds to minimize execution impact slippage.
Study 04: Causal RL in Gymnasium
Adapting legacy RL concepts (`ml_core` EnvFlex / FSMExecution) into an audited Gymnasium environment evaluating multi-objective rewards on double-entry replay ledgers.
Built with Institutional Rigor
Trading algorithms are only as robust as the systems running them. We apply strict software engineering disciplines to financial data infrastructure.
Type Safety
Zero untyped code. Full strict typing across Python 3.12+, Pydantic v2 schemas, and immutable dataclasses with static verification in CI.
Parquet & Arrow
Columnar partitioned storage with Snappy/ZSTD compression. Zero row-by-row parsing overhead; sub-millisecond scans over multi-gigabyte tick datasets.
ADR Architecture
Every contract evolution, fee formula, and gap remediation requires an Architectural Decision Record (ADR) reviewed against frozen baseline specs.
Golden E2E Gates
Adversarial suites (A1–A9), Golden Bybit BTCUSDT end-to-end integration tests, and ledger balance conservation checked on every PR.
