Neural Impulse Predictor: 56-Feature Input Vector Explained

The Neural Impulse Predictor is Whistl's core AI model, processing 56 input features to forecast spending impulses with 84% accuracy. This comprehensive breakdown explains every feature category, how they're measured, and how they combine to predict your financial vulnerability.

The 56-Feature Input Vector

The neural network processes five categories of features:

Category Breakdown

Temporal Features (8 inputs)

Time-based patterns that correlate with impulses:

Feature List

1. Hour of day (0-23): Normalised cyclical encoding
2. Day of week (0-6): Monday=0 through Sunday=6
3. Days since last payday: 0-30 days
4. Days until next payday: 0-30 days
5. Time since last impulse: Hours since last detected impulse
6. Seasonal indicator: Spring/Summer/Autumn/Winter encoding
7. Holiday proximity: Days to nearest public holiday
8. Weekend/weekday flag: Binary indicator

Why Temporal Features Matter

• Circadian rhythms: Willpower fluctuates throughout the day
• Weekly patterns: Fridays/Saturdays often higher risk
• Payday effect: Fresh funds + celebration impulse
• Recovery time: Recent impulses predict near-term risk

Location Features (6 inputs)

Physical proximity to triggers:

Feature List

1. GPS coordinates: Current latitude/longitude
2. Distance to nearest gambling venue: Meters to closest casino/TAB
3. Distance to nearest shopping centre: Meters to closest mall
4. Home/away status: Binary (at home = 1, away = 0)
5. Venue density in area: Number of venues within 2km
6. Location history pattern: Is this a familiar location?

Why Location Features Matter

• Proximity effect: Within 500m increases impulse by 340%
• Environmental cues: Seeing venues triggers associations
• Home safety: Some users more vulnerable at home (boredom)
• Venue density: High-density areas create ambient risk

Biometric Features (8 inputs)

Physiological markers of vulnerability:

Feature List

1. Heart rate variability (HRV): ms, normalised to personal baseline
2. Resting heart rate: bpm, deviation from baseline
3. Sleep duration: Hours slept last night
4. Sleep quality score: 0-100 from sleep tracker
5. Oura readiness score: 0-100 composite recovery score
6. Stress level indicator: From Apple Health/stress apps
7. Activity level: Steps/movement today
8. Recovery status: Combined recovery metrics

Why Biometric Features Matter

• HRV and impulse control: Low HRV = reduced prefrontal function
• Sleep and decision-making: Poor sleep impairs PFC by 40%
• Stress and coping: Elevated stress drives comfort-seeking
• Recovery and resilience: Poor recovery = lower willpower

Financial Features (18 inputs)

Money patterns and current financial state:

Feature List

1. Current account balance: Total available funds
2. Protected floor balance: Reserved for essentials
3. Discretionary balance: Available for non-essentials
4. Spending velocity (7-day): Rate vs. 6-month average
5. Spending velocity (30-day): Rate vs. 6-month average
6. Gambling category ratio: Current month vs. budget
7. Shopping category ratio: Current month vs. budget
8. Dining category ratio: Current month vs. budget
9. Days until overdraft: Projected based on spending rate
10. BNPL active plans count: Number of active Buy Now Pay Later plans
11. Crypto holdings volatility: 7-day price change
12. Recent large transactions: Count of transactions >$500 in 7 days
13. Subscription renewal dates: Days to next renewal
14. Bill payment deadlines: Days to next bill due
15. Savings goal progress: Percentage of goal achieved
16. Credit utilisation ratio: Credit used / credit available
17. Cash withdrawal frequency: ATM withdrawals in 7 days
18. Online transaction ratio: Percentage of transactions online

Why Financial Features Matter

• Velocity indicates失控: Accelerating spending = loss of control
• Budget ratios show pressure: Over-budget categories signal risk
• Overdraft proximity: Financial stress drives coping behaviours
• BNPL stacking: Multiple plans indicate financial stress

Behavioural & Context Features (16 inputs)

Real-time context and behavioural patterns:

Feature List

1. Calendar stress events: Count of stressful events in next 48 hours
2. Screen time patterns: Hours today vs. average
3. Social media usage spikes: Time on social apps vs. baseline
4. App session duration: Current session length
5. DNS gambling queries: Count in last hour
6. DNS shopping queries: Count in last hour
7. Weather conditions: Rainy/sunny/cold/hot encoding
8. Mood check-in score: Self-reported mood (1-10)
9. Journal sentiment: AI-analysed sentiment from entries
10. Recent intervention history: Interventions in last 24 hours
11. Bypass attempt frequency: Bypass attempts in last 24 hours
12. Partner interaction status: Last contact with accountability partner
13. Goal engagement level: Days since last goal check-in
14. Alternative action success rate: Recent success with alternatives
15. Cooldown timer compliance: Percentage of timers completed
16. Self-reported urge intensity: Current urge rating (1-10)

Why Behavioural Features Matter

• Calendar stress: Upcoming deadlines trigger coping
• Screen time: Increased usage correlates with impulsivity
• DNS queries: Active browsing precedes purchases
• Mood and sentiment: Negative emotions drive spending

Feature Normalisation

All features are normalised to 0.0-1.0 scale before input:

Normalisation Methods

# Different normalisation approaches

# Min-Max scaling (for bounded features)
normalised = (value - min) / (max - min)

# Z-score normalisation (for unbounded features)
normalised = 1 / (1 + exp(-(value - mean) / std))

# Cyclical encoding (for time features)
hour_sin = sin(2π × hour / 24)
hour_cos = cos(2π × hour / 24)

# Personal baseline normalisation (for biometrics)
normalised = 1.0 - (current / baseline)
# Example: HRV 35ms vs. baseline 50ms = 0.3 (below baseline)

Feature Importance Ranking

Not all features contribute equally to predictions:

Top 10 Most Predictive Features

Feature Interaction Effects

Features don't act in isolation—they interact:

Key Interactions

• Venue × Time: Being near a venue at night is worse than during day
• Sleep × Stress: Poor sleep amplifies stress effects
• Payday × Velocity: Payday + high velocity = critical risk
• Mood × DNS: Bad mood + browsing = high impulse probability

Privacy: On-Device Feature Processing

All 56 features are processed on your device:

• Location data: Never leaves your phone
• Biometric data: Stays in HealthKit/Oura secure storage
• Financial data: Processed locally after secure bank sync
• Behavioural data: Stored encrypted on-device only

Conclusion

The 56-feature input vector creates a comprehensive picture of your impulse vulnerability. From neural predictions to venue proximity, biometrics to browsing patterns—every feature contributes to life-saving prediction at the moment that matters most.

This isn't just data—it's your personal vulnerability profile, constantly updated, constantly protecting.

Related: AI Predicts Spending Impulses | 27 Risk Signals | On-Device AI Privacy