A study found that an artificial-intelligence model gave the most accurate forecasts when it used material prices and wider economic indicators.
Researchers developed a framework for forecasting quarterly prices across detailed categories of construction materials. It used price histories for each six-digit section in the Construction Specifications Institute’s MasterFormat, along with information such as raw-material prices, producer-price indexes, economic measures and local construction-cost indexes.
The study compared four approaches: a long short-term memory model, ARIMA, a vector error correction model and Chronos-Bolt. The long short-term memory model generally produced the most accurate forecasts, including for representative wood, plastic and metal sections. The researchers tested the framework across 22 CSI divisions, with Division 06—wood, plastics and composites—used as the main detailed example.
How the forecasts performed
Including outside information about material prices and economic conditions improved predictive performance across all four models compared with using CSI price data alone. The long short-term memory model performed best overall.
For four representative six-digit sections, its forecasts had R² values from 0.92 to 0.98, root mean squared errors from 2.26 to 8.70, and mean absolute percentage errors from 0.96 to 2.01. In the study’s reported results, the model reached RMSE as low as 1.390 and MAPE as low as 0.957, and its accuracy improved by up to 59% compared with ARIMA. ARIMA was less accurate but generally followed broad price trends. The vector error correction model and Chronos-Bolt performed poorly across the four sections examined in detail.
Why detailed forecasts matter
Detailed price forecasts could help owners and contractors refine budgets and cost estimates at the stage when project designs and material choices are being developed. Forecasts at the six-digit CSI section level are intended to match the detail needed for definitive estimates.
The framework could also provide material-cost scenarios for financial planning and other analyses, although the study says it does not account for contractor overhead, profit or competitive factors involved in bid estimates.
Evidence and caveats
This was a model-comparison study using quarterly RSMeans material-price indexes and held-out data, rather than a test of whether the forecasts improved outcomes on real construction projects. The researchers evaluated four forecasting approaches, with each model fitted separately to individual six-digit material sections and tested with and without explanatory variables.
Performance varied by section. In Division 06, the reported median RMSE was 5.94 and median R² was 0.96, but some sections had RMSE values as high as 27 and R² as low as 0.55. The researchers attributed weaker results to inconsistent units, unusual price jumps, mixed products and short historical samples. The data covered U.S. residential materials, so the findings may not apply in the same way to commercial construction, other countries or materials outside the study’s coverage.
// Source
Journal of Construction Engineering and Management · 2026 · DOI: 10.1061/jcemd4.coeng-18794
Authors: Iris D. Tommelein, Karnamohit Ranka, Karthik Yeluripati, Junzhe Shi
Institutions: University of California, Berkeley, Baruch College
