Model Architecture

The architecture of a deal scoring model is a weighted probability function. Each input contributes a score. The weights are not set by opinion — they are trained on historical data. Take two hundred closed deals and two hundred lost deals. Feed the five inputs for each deal at the point of scoring. The model learns which inputs mattered most and assigns weights accordingly. In one organization, stakeholder count had a weight of 0.31. In another, meeting frequency was 0.38. The weights are organization-specific because every sales process is different. A model trained on someone else's data will produce someone else's predictions.

1. Step 1: Historical Data Extraction Pull deal records for the last four quarters. Include both won and lost deals. For each deal, capture the five core inputs at the time they reached each stage. You need both outcomes for the model to learn what winning and losing look like.
2. Step 2: Weight Calibration Use logistic regression or a gradient-boosted model to determine the contribution of each input to deal outcome. The output is a weight for each input that reflects its predictive power in your specific sales process. Review the weights. If they do not make operational sense, investigate the data quality.
3. Step 3: Score Calibration Map the raw model output to a 0-100 score. Validate that a score of 80 actually corresponds to an 80% historical close rate. If the calibration is off, adjust. A score that does not correspond to actual probability is a number, not a prediction.