Advanced Trading Algorithms | Quantum Order Flow Models

The Algorithm Stack

Each model serves a specific purpose. Together, they form a complete trading intelligence system that no human trader can replicate manually.

MODEL 01

Delta Z-Score ArbLIVE

Statistical mean-reversion model that calculates the z-score of cumulative delta over a 30-bar window. When |z| ≥ 1.0, the model detects statistically significant order-flow imbalance that historically precedes a directional move.

Window30 bars

Trigger|z| ≥ 1.0

Hold Time15 minutes

OOS PF1.109

MODEL 02

Buyers/Sellers FilterLIVE

Pressure confirmation gate that requires the buyers/sellers imbalance to agree with the z-score sign. Blocks 42% of false z-triggers. The filter alone improved profit factor from 1.11 to 1.58 — a 42% improvement in risk-adjusted returns.

Imbalance Window10 bars>/div>

Confirm Threshold±0.06

Veto Rate42%

OOS PF1.582

MODEL 03

Quantum Signal ProcessorTESTING

Quantum-inspired feature extraction using superposition principles to evaluate multiple market regimes simultaneously. Currently under walk-forward validation. Failed initial gates (AUC 0.4987 vs placebo 0.5008) — will not be deployed until it earns its edge.

RegimesMulti-state

Walk-Forward AUC0.4987

Placebo AUC0.5008

StatusDISABLED

MODEL 04

CVD Slope MomentumLIVE

Measures the rate of change in Cumulative Volume Delta over a 10-bar window. A steep positive slope indicates sustained buying pressure; negative slope indicates distribution. Used as a secondary confirmation for the pressure filter.

Slope Window10 bars

InputCVD delta

Use CaseConfirmation

OOS PF1.348

Validation Methodology

We don't believe in backtests. Every model must survive these gates before it touches live data.

1. Chronological Train/Test Split

Data is split chronologically, not randomly. The first 60% trains the model. The remaining 40% is never seen during training — this is the true out-of-sample test.

2. Walk-Forward Cross-Validation

The model is tested on multiple rolling windows, expanding the training set each time. If performance degrades as the window expands, the model is capturing noise, not signal.

3. Random-Sign Control

We flip the signal direction randomly for a control group. If the random control performs similarly to the model, the model has no edge. The pressure filter beat random control by +488 pips.

4. Cost Stress Test

Models are tested at 2x and 3x realistic spread+commission costs. A model that only works at 1x cost is not production-ready. Pressure filter survived 3x cost (+298 pips).

5. Regime Attribution

Every trade is tagged by market regime (trending, ranging, high/low volatility). If a model only works in one regime, we know when to trust it and when to ignore it.

The Math Behind the Magic

No black boxes. Here's the actual formula that drives the delta z-score model.

# Delta Z-Score Calculation (simplified)
def z_score(delta_series, window=30):
    mean = sum(delta_series[-window:]) / window
    variance = sum((d - mean) ** 2 for d in delta_series[-window:]) / window
    std = variance ** 0.5
    return (delta_series[-1] - mean) / std

# Pressure Confirmation
def confirms(pressure, side, threshold=0.06):
    imb = pressure['imbalance']  # (buy - sell) / (buy + sell)
    return (side * imb) >= threshold  # side: +1 buy, -1 sell

# Signal Logic (the actual production code)
if abs(z) >= 1.0 and confirms(pressure, sign(z)):
    signal = "BUY" if z > 0 else "SELL"
else:
    signal = "HOLD"

See the Algorithms in Action

Watch real-time signals generated by the same models described above. Every signal includes the z-score, pressure confirmation, and exact blocked reason when the filter vetoes a trigger.

Launch the Live Terminal

Algorithms That SeeWhat Humans Can't.