Atul Iwale · Project 13 · Machine learning

Cost Plan Estimator

Estimate the construction cost of an Indian building project at feasibility stage, with a realistic range. The model learned from 1,800 completed projects across 15 cities, and was tested only on projects that started after its training data ended.

Ridge + XGBoost ensembleConformal P10–P90 range Linear · Decision tree · Random forest · Neural network comparedSynthetic data, patterned on real cost plans

Try an estimate

Change any input. The estimate, range and cost drivers update instantly, computed in your browser from the exported model.

How the models compare

Mean absolute percentage error on unseen projects that started in 2024–2025. The coefficient method (base rate × specification × city) is the traditional baseline every model had to beat.

Accuracy by building type

Model error versus the coefficient method on the 2024–2025 test projects.

What drives the estimate

Top XGBoost features by gain: how much each one improves the model's splits.

Portfolio check: are my live contract values realistic?

The 21 building projects from Project 01 (Cost & Margin Intelligence), estimated by the model and compared with their contract values. Projects outside the P10–P90 range are flagged for commercial review, not treated as errors.

How it works

  • Data cleaned first: 6 duplicates removed, 18 areas entered in sq ft corrected, legacy city names and inconsistent types standardised, missing values kept as explicit categories.
  • Costs rebased to January-2018 prices with a composite steel, cement and labour index, the way quantity surveyors compare projects across years.
  • Time-based split: trained on 2018–2023 starts, tested on 2024–2025 starts only.
  • Seven models tuned by 5-fold cross-validation; the best is an average of Ridge regression and XGBoost.
  • P10–P90 range from XGBoost quantile models, calibrated on 2023 projects (conformal calibration) so it covers about 80% of outcomes.

Honest limits

  • The data is synthetic, generated to behave like real cost data (noise, outliers, missing fields). Real-world accuracy must be measured on a client's own completed projects.
  • Costs exclude land, fees and GST. The cost-index weights (25% steel, 20% cement, 40% labour, 15% other) are an assumption.
  • Start months after December 2025 use the latest published index, so future escalation is not forecast.
  • An early-stage estimate, not a substitute for a quantity-based cost plan.

Full workings: Jupyter notebook · README · dataset