SYSTEM ARCHITECTURE & METHODOLOGY
The Quantitative Thesis
Breakout setups are verified through a two-stage statistical filter, and the rules are re-calibrated every 30 days to protect your capital.
Two-Filter Trade Verification Engine
Anyone can catch a breakout. The filter decides whether it is worth your capital. Two independent checks run per setup: the price pattern, then a volatility and order-flow gate.
Deployment week: three candidates fired across 300+ bars. The gate scored all three negative-EV and blocked every one. Zero trades was the correct output.
Source: Master System Report, Section 3 (expectation bands and the certified 3R judge distribution)
Rolling Walk-Forward Framework
Markets change regime. The model re-fits monthly and never scores itself on data it has already seen.
| Parameter | Setting | Why |
|---|---|---|
| Training window | 9 months rolling | Walk-forward optimization puts the stable ridge at 30 to 42 weeks, peaking at 38. |
| Refit cadence | First weekend monthly | Sub-monthly and weekly refits were tested and rejected: higher variance, fitted to transient noise. |
| Folds evaluated | 32 | Forward-only scoring under a strict time embargo, never on data the model has seen. |
| Folds positive | 29 of 32 | All 2025 and 2026 folds green. |
| Ensemble size | 300 trees | Per calibrated model set. |
| Factor streams | 14 | Precious metals, equity indices and FX on M5 and M15, decorrelated so capital survives divergent regimes. |
Source: Master System Report, Section 1 (monthly calibration schedule and risk policy)
Why This Is Not Curve-Fitted
Every in-sample backtest looks immaculate; that is what overfitting means. The deployed book is not validated by its backtest. It is validated by four structural facts, each checkable in the research library.
- The edge survives data the model never saw. Scoring is forward-only: rolling 9-month trainings, then out-of-sample evaluation under a strict time embargo. A curve-fit collapses the moment it meets unseen data; this did not.
- The gate carries the edge, not the fit. The raw breakout pool scores +0.088R un-gated; the deployed, ML-gated book scores +0.297R on identical folds. The alpha lives in the trade-selection filter, which is re-trained monthly, not in one frozen parameter set.
- Regime collapse is a tripwire, not a hope. If volatility regimes shift under the model, the drift tripwires halt dispatch: a month below -15% forces recalibration; two consecutive bottom-percentile months kill the build until it recertifies. Degraded models stop trading before they stop winning.
- Every deployment clears the ruin gate. Each candidate configuration runs 10,000 randomized bootstrap paths with the full risk stack active. Zero ruin events is a pass condition, not a statistic: any nonzero ruin probability blocks the deployment before it ships.
Sources: Production Architecture & Empirical Benchmarks (folds, gate delta, ruin gate) and Loss Forensics (how the losing distribution is priced, not smoothed).
Trading stops on a broken model, never on a bad month.
Both tripwires were fixed in writing before the first trade and are machine-enforced. Drawdown depth alone never halts a system: that rule was tested and it truncates the recovery cycle.
Operational Drift Tripwires
Three standing rules, each with the condition it answers and the action it takes.
| Tripwire | Condition | Action |
|---|---|---|
| T1 soft | Single-month return below -15% | Immediate recalibration and feature verification before further order dispatch. |
| T2 hard | Two consecutive months below p5, or any anchor model failure | Complete entry halt until full recertification passes under the Model Integrity Protocol. |
| Standing prohibition | Drawdown depth alone | Never halts entry. Halting a system on temporary drawdown depth was empirically proven to truncate the recovery cycle. |
Source: Master System Report, Section 3 (drift triggers and Section 8 standing prohibitions)
Code-to-Broker Verification
Nothing reaches an account until it passes automated parity checks and 10,000 randomized stress paths. Six gates, all enforced on every build.
- Machine learning bit-parity. Zero-drift mathematical equivalence between the offline Python feature pipelines and the production execution trees, including bit-for-bit ML scoring.
- Glide curve enforcement. Scale-dependent reverse-exponential risk throttling verified across every account equity tier.
- Execution physics check. Strict validation of broker contract specifications, minimum lot quantization and spread limits.
- Autopsy consistency. Post-trade analytics must match simulation assumptions. The gate caught four backtest winners that were stop-placement artifacts, exposing -76R of hidden losses.
- Governor bounds. Hardcoded exposure limits and maximum simultaneous position caps across correlated baskets.
- Variable map certification. SHA-256 validation of the compiled model sets and model tree count parameters.
Source: Parity Certification Report
Canonical sourced documents
Every quantitative statement and statistical parameter on this page is derived directly from audited research documentation.
- Production Architecture Manual SECTION 3 Recalibration rhythm, multi-factor validation, expectation bands and drift triggers.
- Certified Risk Glide SECTION 1 Certified 14-stream multi-factor model parameters, risk bounds and validation rules.
- Loss Forensics PAPER 04 How the losing distribution is priced rather than smoothed, with four named failure modes.
- The public tape 50 DEALS The audited ledger this filter produced, with the losing trades attached.
- Risk and scaling 50 UNITS The reverse-exponential glide that keeps the same filter honest as capital grows.
- The execution playbook START How a verified setup becomes a scheduled, bracketed order on your account.
- Capacity and pricing 1 CHANNEL The Telegram seat ladder that keeps the pool at a size the edge survives; the planned Stream API remains non-purchasable.