The Lab

The paper phase is over. This is live capital execution.

The gauntlet every method had to survive — research validation, purged walk-forward testing, end-to-end paper operation — has been run. The swarm now executes real capital through institutional pipelines on Google Cloud: an ELO-ranked strategy pool, pre-trade risk checks, regime-aware adaptation, and TWAP/VWAP execution against broker gateways. Operated, monitored, and improved by the same engineer you'd be working with.

Research — validated or killed

Purged walk-forward testing, deflated Sharpe, overfitting probability, regime-segmented analysis. Most ideas died here. That was the point.

Paper — production conditions end-to-end

Full stack in production: ingestion queues, in-memory hot tier, warden gates, execution routing — against live market data, simulated fills, real failure drills. No method touched capital before surviving this.

Live capital — current phase

Real orders, real fills, real drawdown. The same gates, now with money attached: dynamic drawdown thresholds, real-time position sizing, automated kill-switch fail-safes. This phase never ends — the tournament keeps running and capital follows ELO rank.

The Swarm — Live Capital Deployment

Live · GCP · broker gateways

A distributed multi-agent ecosystem where strategy genomes earn capital through ELO tournament play. The Evolution Engine mutates and ranks; the Risk Warden gates every order with seven pre-trade checks, dynamic drawdown thresholds, real-time position sizing, and automated kill-switch fail-safes; the Execution Layer routes optimal TWAP/VWAP schedules through broker gateways; and multi-state regime adaptation (Bull / Bear / Neutral / Macro Volatility) re-weights the book in real time.

8 genomesELO tournament-ranked, live allocation

7 gatesPre-trade checks on every order

4 regimesRegime posterior drives posture & sizing

c# .net 8/9 python cloud run spot vm batch memorystore redis versioned gcs

Read how the engine plans & learns →

Live

Balanced Portfolio Engine

The conservative book: regime-aware sizing, portfolio-level exposure caps, and drawdown circuit breakers that de-risk automatically. Now keyed off the swarm's shared regime posterior rather than its own detector.

regime detection risk parity circuit breakers

Live

Frontier Optimizer

Efficient-frontier construction fused with an NLP news-sentiment pipeline that tilts allocations as the information environment shifts. Mathematics allocates; language models read the room.

mean–variance sentiment nlp cloud build

Live

Volume Predictor

ML intraday volume forecasting — the execution-timing signal that feeds TWAP/VWAP scheduling across the swarm, and the first candidate for the FluxMetrics public API.

ml forecasting microstructure api-ready

Signal precision & risk-adjusted telemetry

Monitored continuously

Every candidate signal passes a stacked-ensemble conviction filter before it can become an order: the meta-learner scores conviction across breakout, duration, return, and risk-reward sub-models and drops low-conviction noise below the trigger floor. Risk-adjusted performance is tracked with multiple-testing-corrected statistics — Deflated Sharpe Ratios computed on both the historical record and the live-capital book — alongside max drawdown containment and regime-segmented performance matrices.

65%+ Signal win probability filtered ensemble output, engineered floor

P95 Ensemble conviction floor low-conviction triggers dropped pre-trade

DSRΔ Deflated Sharpe, live tracking historical + live-capital series

<1ms Hot-path state access memorystore redis · async ingestion

These figures describe engineered system characteristics and internal monitoring targets, not promised returns. Verified production metrics (live DSR series, drawdown containment, per-regime attribution) are disclosed to qualified institutional counterparties under NDA — request institutional access.

Backtest vs. live — what changed at each gate

The transition to live capital didn't loosen the harness; it tightened it. What each phase measures, and what changed when money became real:

Dimension	Research / paper phase	Live capital phase
Fill model	simulated (spread + slippage model)	real fills, reconciled vs. broker state
Risk gating	warden gates in shadow mode	7-gate hard enforcement · kill-switch armed
Execution	idealized bar-close assumptions	TWAP/VWAP schedules · impact-aware child orders
Performance stats	deflated Sharpe (historical)	deflated Sharpe (historical + live series)
Failure drills	chaos-tested in paper	live: DLQs, circuit breakers, drawdown thresholds
Capital	paper	live — firm's own capital

Regime performance matrix

The book is segmented by the regime posterior — the same four states that drive sizing and posture. Structure below; per-regime verified metrics ship under NDA:

Regime	Posture	Preferred genomes	Drawdown posture
Bull	risk-on, trend-following weights up	dual-ema · adx-expansion · tmfc	trailing stops loose, trail = f(ATR)
Bear	short-bias / defensive weights up	rsi-snapback · bollinger-reversion	tighter trail, faster STAND_DOWN trigger
Neutral	mean-reversion weights up, size trimmed	stoch-flow · bollinger-reversion	standard ladder, multi-tier TP
Macro Volatility	risk-off — STAND_DOWN dominant	volume-spike (liquidity watch only)	de-risk cascade → kill-switch armed

Regime segmentation is also how we keep ourselves honest: performance is reported per-state, so a strategy that only works in one regime can't hide behind the average.

The strategy library — eight genome-optimized deployments

Each strategy ships as an evolving genome: parameters are mutated, tournament-tested on identical market paths, ELO-ranked, and only then trusted with capital. Descriptions below are capability-level; genome files and hyperparameters stay in the vault.

STRAT / 01

Dynamic RSI Snap-Back

Regime-conditioned overbought/oversold oscillators with adaptive threshold elasticity. Entries only when the regime posterior confirms the mean-reversion window; threshold bands breathe with realized volatility.

STRAT / 02

Dual EMA Crossover

Trend-following momentum with dynamic filter bands. EMA pair widths and confirmation windows are genome-controlled and re-ranked per regime by the tournament engine.

STRAT / 03

MACD Volatility-Scaled

Adaptive signal-line velocity and histogram divergence, scaled by realized volatility so the same genome stays calibrated across quiet and violent tapes.

STRAT / 04

ADX Trend Expansion

Regime-gated directional index tuned for explosive trend capture. Stands down unless expansion conditions and the regime posterior both agree.

STRAT / 05

Volume Breakout Spike

Liquidity surge detection coupled with institutional footprint tracking — volume spikes are validated against flow footprint before the swarm treats them as breakouts.

STRAT / 06

Dynamic Bollinger Mean-Reversion

Volatility-envelope reversion with multi-tier take-profit ladders, sized against live drawdown budgets and gated by warden headroom.

STRAT / 07

Stochastic Momentum Flow

Fast/slow %K/%D momentum cycles calibrated across multi-timeframe panels for cycle-turn timing — panel weights are genome parameters, not defaults.

STRAT / 08

TMFC Composite (Trend–Momentum–Flow)

Multi-factor composite fusing price action, momentum, and cross-asset flow signals into a single conviction score — the swarm's tie-breaker book.

Interactive: regime detection, live in your browser

A year of prices is simulated from a hidden multi-state regime model (bull / chop / crisis). Press run and a recursive Bayesian filter — the same recursion inside our regime adaptation layer — recovers the hidden regime from daily returns alone, one day at a time, no look-ahead.

bull (µ>0, low σ) chop (µ≈0, mid σ) crisis (µ<0, high σ) de-risk zone: P(crisis)>50%

α_t(j) ∝ 𝒩(r_t | µ_j, σ_j) · Σ_i α_t−1(i) P_ij — hover the chart after the sweep to read the posterior on any day.

A note on honesty

The gauntlet is complete: research → paper → live capital. We publish what each system does and how it's engineered, not cherry-picked return figures. Verified live metrics — deflated-Sharpe series, drawdown containment, regime matrices — are disclosed to qualified institutional counterparties under NDA. Request institutional access or book a working session and we'll walk you through the dashboards.

Risk disclosure: The Lab documents engineering and operational practice, not investment advice or an offer of securities. Trading involves substantial risk of loss; past or hypothetical performance — simulated or live — is not indicative of future results. System metrics on this page describe engineered characteristics and internal monitoring targets.

Want this discipline applied to your stack?

We'll build yours the way we run ours.

Strategy validation, data pipelines, execution infrastructure, risk systems — engineered, documented, and handed over. Start with a free 30-minute working session.

Schedule a session Institutional access