Model Selection

The best model is the simplest one that meets your accuracy requirements. Not the most sophisticated. Not the one with the most impressive name. The simplest one that works.

I have watched data teams spend six months building neural networks that outperform logistic regression by 3% on a churn prediction task. Three percent. At a cost of 10x development time, 5x maintenance burden, and zero interpretability. The logistic regression would have shipped in two weeks and told you exactly which factors drive churn. The neural network is a black box that is 3% more accurate and infinitely harder to explain to the VP who needs to act on it.

MODEL SELECTION GUIDE
======================

TASK                → START WITH            → UPGRADE IF NEEDED
────────────────────┬──────────────────────┬─────────────────────
Churn prediction    │ Logistic regression  │ Gradient boosting
Lead scoring        │ Logistic regression  │ Random forest
Revenue forecast    │ Time series (ARIMA)  │ Prophet / LSTM
Demand prediction   │ Linear regression    │ Gradient boosting
Anomaly detection   │ Statistical (Z-score)│ Isolation forest
Classification      │ Decision tree        │ Gradient boosting
Clustering          │ K-means              │ DBSCAN

SELECTION CRITERIA:
  1. Interpretability required?
     Yes → Logistic regression, decision tree, linear
     No  → Gradient boosting, neural network, ensemble

  2. Data volume?
     < 1,000 rows → Simple models only (complex overfit)
     1,000-10,000 → Standard ML models
     > 100,000    → Deep learning becomes viable

  3. Maintenance budget?
     Low  → Simple, retrain quarterly
     High → Complex, retrain monthly, monitor drift

THE RULE: Start simple. Upgrade only when accuracy gains
justify the complexity cost. Track both accuracy AND
maintenance burden.

AI has changed this calculus. General-purpose AI models can now approximate specialized ML models for many business prediction tasks. Feed historical deal data to an AI with the prompt "predict which deals will close this quarter and explain why," and you get a prediction with reasoning that a purpose-built ML model would have taken weeks to develop.

The accuracy gap: purpose-built ML models outperform general AI by 10-20% on structured prediction tasks. But the development gap is 2 weeks versus 2 months. For 80% of business decisions, the AI approximation is sufficient. For the 20% where every accuracy point matters — fraud detection, medical diagnosis, financial trading — build the specialized model.