Optimize Lead Scoring with ML and Rule-Based Models

This skill covers designing and implementing lead-scoring models for sales and marketing - both rule-based and machine-learning approaches, covering business context, data requirements, model evaluation, and practical implementation challenges. Scoring-framework design: explicit (demographic/firmographic) versus implicit (behavioral) scoring balance, positive and negative scoring (engagement indicators plus disqualifying factors), time-based decay functions for behavioral signals, score normalization to a consistent 0-100 scale, and multi-dimensional scoring separating fit (ICP alignment) from intent (buying signals). Data foundation: lead demographics (title, seniority, department, company size, industry, geography), behavioral data (email opens/clicks, website visits, content downloads, webinar attendance), engagement patterns (frequency, recency, depth), and historical sales-outcome data for training and validation.

Rule-based scoring is demonstrated via a scoring configuration and function:

### Rule-based lead scoring configuration
SCORING_RULES = {
    'demographic': {
        'job_title': {
            'C-Level': 25, 'VP': 20, 'Director': 15, 'Manager': 10, 'Individual Contributor': 5
        },
        'company_size': {
            '1000+': 20, '500-999': 15, '100-499': 10, '50-99': 5, '<50': 0
        },
        'industry_fit': {
            'high_fit': 20, 'medium_fit': 10, 'low_fit': 0, 'poor_fit': -10
        }
    },
    'behavioral': {
        'email_engagement': {'open': 2, 'click': 5, 'reply': 10},
        'website_activity': {'visit': 3, 'multiple_pages': 7, 'pricing_page': 15, 'demo_request': 25},
        'content_engagement': {'download': 8, 'webinar_attendance': 12, 'trial_signup': 30}
    },
    'negative_indicators': {
        'competitor': -50, 'student_email': -20, 'out_of_territory': -30
    }
}

def calculate_lead_score(lead_data):
    score = 0
    
    # Demographic scoring
    for category, rules in SCORING_RULES['demographic'].items():
        if lead_data.get(category) in rules:
            score += rules[lead_data[category]]
    
    # Behavioral scoring with recency decay
    for activity in lead_data.get('activities', []):
        activity_type = activity['type']
        days_ago = activity['days_ago']
        
        base_score = SCORING_RULES['behavioral'].get(activity_type, {}).get(activity['action'], 0)
        # Apply decay: 100% for 0-7 days, 75% for 8-30 days, 50% for 31-90 days
        if days_ago <= 7:
            decay_factor = 1.0
        elif days_ago <= 30:
            decay_factor = 0.75
        elif days_ago <= 90:
            decay_factor = 0.5
        else:
            decay_factor = 0.25
            
        score += base_score * decay_factor
    
    # Apply negative indicators
    for indicator, penalty in SCORING_RULES['negative_indicators'].items():
        if lead_data.get(indicator, False):
            score += penalty
    
    return max(0, min(100, score))  # Normalize to 0-100

The machine-learning approach covers feature engineering (seniority score, log-transformed company size, industry-fit flag, email-engagement rate, page-view and session-duration features, recency features, an engagement-velocity interaction feature) and a training pipeline using a gradient-boosting classifier on a stratified train/test split with scaled features, converting predicted probabilities to a 0-100 score.

Model evaluation covers precision, recall, F1, and AUC plus business metrics - conversion rate among high-scoring leads and the resulting lift over the overall conversion rate. Implementation best practices define score-threshold tiers (hot 80-100: immediate sales outreach; warm 60-79: marketing nurture with sales notification; cold 40-59: automated email sequences; unqualified under 40: minimal touch, education-focused) and data-quality validation (required-field checks, email-format validation, spam-domain detection for the company field). An A/B testing framework randomly assigns leads to test groups, applies different scoring models per group, and tracks conversions by group. Integration and deployment cover a CRM sync pattern (batch-updating lead scores in 200-record batches with a last-updated timestamp and score-breakdown reason) and a real-time scoring API - a web endpoint validating input, calculating the score, mapping it to a priority tier, and returning a recommended action and score breakdown.

Use it when designing, building, or evaluating a lead-scoring model - a rule-based point system, a machine-learning classifier, threshold-based routing tiers, or CRM/API integration for scores. It is not a full marketing-automation or CRM platform - it produces the scoring logic and thresholds that feed those systems, not the systems themselves.

Given lead demographic and behavioral data, and for the ML approach historical conversion outcomes, it produces a 0-100 lead score with a priority tier, model-evaluation metrics (precision, recall, AUC, lift), a data-quality validation result, and CRM-sync or API-response payloads.

Code samples use pandas/numpy for feature engineering, scikit-learn (GradientBoostingClassifier, StandardScaler, train_test_split, evaluation metrics) for the ML model, and Flask for a real-time scoring API integrated with a CRM client via batch updates.

Marketing operations, sales operations, and revenue teams building or maintaining a lead-scoring model.