Development of an AI system for predicting deals in the sales funnel
Sales forecasting is a key planning task. Manual forecasts by managers are systematically biased: overoptimistic and anchored on wishful thinking. The ML model is based on objective signals from CRM and closing history.
Forecasting model
Deal Level: For each active deal, the probability of closing in the current month/quarter is calculated. XGBoost uses the following features: pipeline stage, amount, time in stage, number of contacts, last activity, progression rate, and manager historical data.
Aggregate Forecast: A weighted sum of probabilities across all trades. Confidence interval (p10–p90) is more accurate than a point forecast. Comparison with pipeline history.
Cohort Analysis: Deals created in the same period, for some reason, close worse - early identification through cohort tracking.
Time-series component
Prophet / Temporal Fusion Transformer for accounting for seasonality, trends, external factors (month, quarter-end, holidays).
Dashboard
The real-time forecast is updated with each CRM event. Waterfall chart: pipeline start → new deals → projected wins → projected losses → forecast.
Alert: If the quarterly forecast deviates from the plan by >15%, automatic notification to management.
Data requirements
18–24 months of historical closed trade data for a reliable forecast. Minimum 200 closed trades.