Risk Orchestrator: Composite Score Calculation Explained

The Risk Orchestrator is Whistl's central risk assessment engine. It combines 27 weighted signals into a single composite score (0.0-1.0) that determines when to activate blocking, trigger intervention, and notify partners. This deep dive explains the mathematics behind your protection.

Why a Composite Score?

Individual signals aren't enough—risk emerges from signal combination:

The Limitation of Single Signals

• Venue proximity alone: Being near a casino doesn't mean you'll gamble
• Poor sleep alone: Being tired doesn't mean you'll spend impulsively
• Payday alone: Getting paid doesn't mean you'll blow it

The Power of Combination

• Venue + poor sleep + payday: Now we're talking—triple risk
• Synergistic effects: Combined risk > sum of individual risks
• Context matters: Same signal means different things in different contexts

The Composite Score Formula

The Risk Orchestrator calculates a weighted sum of all 27 signals:

Basic Formula

# Composite risk score calculation
composite_risk = Σ(signal_value × signal_weight)

# Where:
# - signal_value: Current value of each signal (0.0 - 1.0)
# - signal_weight: Importance of each signal (sums to 1.0)
# - composite_risk: Final score (0.0 - 1.0)

# Example calculation with 5 signals (simplified):
signal_1_value = 0.8  # Neural prediction: high
signal_1_weight = 0.127  # 12.7% weight

signal_2_value = 0.6  # Spending velocity: moderate
signal_2_weight = 0.118  # 11.8% weight

signal_3_value = 1.0  # Venue proximity: at venue
signal_3_weight = 0.059  # 5.9% weight

# Partial sum:
partial_risk = (0.8 × 0.127) + (0.6 × 0.118) + (1.0 × 0.059)
             = 0.1016 + 0.0708 + 0.059
             = 0.2314

# Add remaining 24 signals...
# Final composite_risk = 0.73 (HIGH)

Signal Value Normalisation

All signals are normalised to 0.0-1.0 scale:

# Normalisation examples

# HRV: Lower is worse (inverse relationship)
hrv_value = 1.0 - (current_hrv / baseline_hrv)
# If baseline = 50ms, current = 35ms
# hrv_value = 1.0 - (35/50) = 1.0 - 0.7 = 0.3

# Sleep: More is better (up to optimal)
if sleep_hours < 7:
    sleep_value = (7 - sleep_hours) / 7  # Deficit
elif sleep_hours > 9:
    sleep_value = (sleep_hours - 9) / 3  # Excess
else:
    sleep_value = 0  # Optimal range

# Distance to venue: Closer is worse
if distance < 100:  # meters
    venue_value = 1.0  # At venue
elif distance < 500:
    venue_value = 0.8  # Very close
elif distance < 1000:
    venue_value = 0.5  # Close
elif distance < 2000:
    venue_value = 0.2  # Nearby
else:
    venue_value = 0  # Safe distance

Signal Weights by Tier

The 27 signals are organised into tiers based on predictive power:

Tier 1: Primary Predictors (>10% weight each)

Tier 1 Total: 34.3% of composite score

Tier 2: Strong Predictors (5-10% weight each)

Tier 2 Total: 23.6% of composite score

Tier 3: Moderate Predictors (2-5% weight each)

Tier 4: Contextual Predictors (1-2% weight each)

Risk Thresholds and Actions

The composite score determines protective actions:

Threshold Table

Hysteresis for Stability

To prevent rapid state flipping, thresholds have hysteresis:

# Hysteresis implementation
def should_change_state(current_state, new_score):
    if current_state == "GREEN":
        # Must exceed 0.45 to leave GREEN (not just 0.40)
        return new_score > 0.45
    elif current_state == "YELLOW":
        # Must drop below 0.35 to go to GREEN
        # Must exceed 0.65 to go to ORANGE
        if new_score < 0.35:
            return "GREEN"
        elif new_score > 0.65:
            return "ORANGE"
        return "YELLOW"
    elif current_state == "ORANGE":
        # Must drop below 0.55 to go to YELLOW
        # Must exceed 0.85 to go to RED
        if new_score < 0.55:
            return "YELLOW"
        elif new_score > 0.85:
            return "RED"
        return "ORANGE"
    elif current_state == "RED":
        # Must drop below 0.75 to leave RED
        return new_score < 0.75

Real-World Calculation Example

Marcus, Friday 8:30pm, near Crown Casino:

Signal Values and Contributions

State Determination

Composite score of 0.508 = YELLOW state (Elevated risk)

Actions: Proactive check-in, increased monitoring, partner notification optional

Weight Adaptation Over Time

Signal weights aren't static—they adapt based on predictive accuracy:

Exponential Moving Average Updates

# Weight adaptation algorithm
def update_signal_weight(signal_name, current_weight, prediction, outcome):
    learning_rate = 0.05  # How quickly weights adapt
    
    # Calculate prediction error
    error = abs(outcome - prediction)
    
    # If signal predicted correctly, increase weight
    if error < 0.1:
        # Reinforce: move weight toward maximum (0.20)
        adjustment = learning_rate * (1.0 - current_weight)
        new_weight = current_weight + adjustment
    # If signal predicted incorrectly, decrease weight
    elif error > 0.3:
        # Reduce: move weight toward minimum (0.01)
        adjustment = learning_rate * current_weight
        new_weight = current_weight - adjustment
    else:
        # Small error: minor adjustment
        new_weight = current_weight
    
    # Ensure weights stay in valid range
    new_weight = max(0.01, min(0.20, new_weight))
    
    return new_weight

Personal Calibration

After 30 days, your weights are uniquely calibrated:

• If venue proximity predicts YOUR impulses well → Weight increases
• If weather doesn't predict YOUR impulses → Weight decreases
• Your model differs from other users based on your patterns

Effectiveness Data

Risk Orchestrator performance metrics:

User Testimonials

"I don't understand all the math, but I know when Whistl says my risk is high, it's right. The composite score thing works." — Marcus, 28

"Love seeing the breakdown of what's driving my risk. Helps me understand my triggers." — Sarah, 34

"The fact that it learns my patterns over time is wild. It knows me better than I know myself." — Jake, 31

Conclusion

The Risk Orchestrator transforms 27 disparate signals into a single, actionable risk score. Through weighted combination, adaptive learning, and intelligent thresholds, it creates a comprehensive picture of your impulse vulnerability.

This isn't just math—it's your personal risk radar, constantly scanning, constantly learning, constantly protecting.

Related: 27 Risk Signals | SpendingShield | Adaptive Weight System