Lead scoring ML et pipeline forecasting

Section 14.4.1 : Lead scoring ML et pipeline forecasting

🎯 Objectif pédagogique

Implémenter un système de lead scoring ML en production et un modèle de forecasting de pipeline commercial pour prédire les revenus futurs avec précision.

Lead scoring ML en production

Le scoring ML va au-delà du prototype — il faut le maintenir et le monitorer en production :

# Pipeline ML complet pour lead scoring en production

# 1. Feature engineering
def create_features(lead_data, enrichment_data, activity_data):
    features = {
        # Firmographiques (enrichissement)
        'company_size': enrichment_data.get('employees', 0),
        'industry_match': 1 if enrichment_data.get('industry') in TARGET_INDUSTRIES else 0,
        'funding_total': enrichment_data.get('funding_total', 0),
        'tech_stack_overlap': calculate_tech_overlap(enrichment_data.get('tech', [])),

        # Démographiques
        'seniority_score': seniority_to_score(lead_data.get('title', '')),
        'department_match': 1 if is_target_department(lead_data.get('title', '')) else 0,

        # Comportementaux (activité)
        'pages_visited_7d': activity_data.get('pages_7d', 0),
        'time_on_site_7d': activity_data.get('time_7d', 0),
        'pricing_page_visits': activity_data.get('pricing_visits', 0),
        'content_downloads': activity_data.get('downloads', 0),
        'email_opens_30d': activity_data.get('email_opens_30d', 0),
        'email_clicks_30d': activity_data.get('email_clicks_30d', 0),

        # Temporels
        'days_since_first_touch': activity_data.get('days_since_first', 0),
        'days_since_last_activity': activity_data.get('days_inactive', 0),
        'activity_acceleration': activity_data.get('acceleration', 0)
    }
    return features

# 2. Scoring en temps réel
def score_lead(lead_id):
    lead = CRM.get_lead(lead_id)
    enrichment = EnrichmentService.get(lead['email'])
    activity = ActivityTracker.get(lead_id)

    features = create_features(lead, enrichment, activity)
    score = model.predict_proba([list(features.values())])[0][1]

    # Mettre à jour le CRM
    CRM.update_lead(lead_id, {
        'ml_score': round(score * 100),
        'ml_score_updated_at': datetime.now(),
        'ml_tier': 'hot' if score > 0.7 else 'warm' if score > 0.3 else 'cold'
    })

    return score

Pipeline forecasting

La prédiction du pipeline permet d'anticiper le revenu futur et de prendre des décisions business informées :

# Forecasting de pipeline avec ML

import pandas as pd
from sklearn.ensemble import RandomForestRegressor

def forecast_pipeline():
    # Données : chaque deal avec ses caractéristiques
    deals = pd.DataFrame([{
        'deal_size': deal.amount,
        'days_in_stage': deal.days_in_current_stage,
        'total_activities': deal.total_activities,
        'champion_engaged': deal.champion_score,
        'competitor_mentioned': deal.competitor_flag,
        'stage_numeric': stage_to_number(deal.stage),
        'historical_win_rate_segment': segment_win_rate(deal),
        'closed_won': deal.is_won  # Target
    } for deal in Deal.objects.filter(closed=True)])

    # Entraîner le modèle
    X = deals.drop('closed_won', axis=1)
    y = deals['closed_won']
    model = RandomForestRegressor(n_estimators=100)
    model.fit(X, y)

    # Prédire sur le pipeline actuel
    open_deals = Deal.objects.filter(closed=False)
    for deal in open_deals:
        features = extract_features(deal)
        win_probability = model.predict([features])[0]
        weighted_amount = deal.amount * win_probability

        deal.ai_win_probability = win_probability
        deal.ai_weighted_amount = weighted_amount
        deal.save()

    # Forecast total
    total_pipeline = sum(d.amount for d in open_deals)
    weighted_pipeline = sum(d.ai_weighted_amount for d in open_deals)

    return {
        'total_pipeline': total_pipeline,
        'weighted_forecast': weighted_pipeline,
        'confidence_interval': calculate_confidence(open_deals)
    }

Dashboard de forecasting