Forecasting Construction Material Prices, One Category at a Time

Construction estimators need more than a prediction that building costs will rise. They need to know which material categories might become more expensive and how those changes could affect a detailed estimate. A study led by Boge Lyu investigates that problem by organizing forecasts around the Construction Specifications Institute’s MasterFormat classification system.

The work is available as a revised arXiv preprint, although the supplied manuscript also records publication in the Journal of Construction Engineering and Management. Its central finding is specific: in the detailed benchmark for Division 06, covering wood, plastics and composites, a long short-term memory neural network, or LSTM, produced the strongest reported forecasts among five tested approaches.

A forecast between the item and the whole project

MasterFormat separates broad, two-digit Divisions from more detailed, six-digit Sections. The researchers use those Sections as forecasting units. Each represents a category-level price signal rather than an individual product quotation or a complete project budget.

Their RSMeans-based dataset contains 11,185 material-bearing items mapped into 455 Sections across 22 Divisions, using publications spanning 2007–2025. Prices are processed quarterly, adjusted using the San Jose City Cost Index and converted into indices for comparison over time. Service-related content is excluded to keep the analysis focused on material-linked costs.

The researchers supplement these histories with 145 producer price index series and economic indicators covering inflation, employment, housing activity, construction spending and commodity markets. The aim is to give models information about upstream prices and broader market conditions, not just a material category’s own past movements.

Comparing five forecasting approaches

The study compares LSTM, ARIMA, a vector error correction model, a vanilla Transformer and Chronos-Bolt, a pretrained time-series forecasting model. Baseline configurations use historical construction-price information; augmented configurations add external indicators. Chronos-Bolt is evaluated in zero-shot, fine-tuned and regressor-assisted variants.

LSTMs use gated memory to learn patterns in sequences. ARIMA models statistical relationships with earlier observations and errors, while vector error correction models represent short-run adjustments around longer-run relationships. Transformer-based approaches use attention mechanisms to learn relationships within sequences.

The reported training protocol reserves the latest 15% of observations for testing. Model selection uses chronological, walk-forward validation rather than randomly mixing earlier and later observations. During sequential testing, forecasts are made before the corresponding outcome is observed, and newly available observations are then used for updates.

That distinction matters: the results should not be read as the accuracy of a single forecast issued years in advance without subsequent observations or model updates.

Where the LSTM performed best

For Division 06, the augmented LSTM achieved a median root mean squared error of 4.19 price-index points, a median Section-level mean absolute percentage error (MAPE) of 1.60% and a median R-squared of 0.96. These are medians across Section-level metrics, not guarantees for every material category.

The augmented ARIMA recorded a median RMSE of 14.82 and a median Section-level MAPE of 6.50%. The vanilla Transformer reached 15.41 and 6.29%, respectively. Both improved substantially over their baselines, as did the LSTM.

However, the detailed results do not support the abstract’s broad claim that external variables improved every model. The augmented vector error correction model performed worse than its baseline. Chronos-Bolt also remained comparatively inaccurate across its tested variants. The evidence therefore favors particular model-and-feature combinations, not adding economic indicators indiscriminately.

The researchers also report that separate models for individual Sections outperformed their pooled LSTM alternative. Feature-attribution analysis identifies contributions from price indices, economic indicators and engineered historical features, but does not establish those variables as causal drivers of prices.

Useful inputs, not proven project savings

The intended application is later-stage estimating, when quantities and specifications are sufficiently developed to use granular cost inputs. Forecasts could inform unit-cost updates, escalation allowances and procurement reviews. The study does not directly demonstrate lower final bids, fewer overruns or better purchasing outcomes.

Important boundaries remain. Some Sections contain mixed-unit inconsistencies, anomalous movements and heterogeneous products. The authors report testing across 22 Divisions, but present Division 06 in detail and caution about generalization beyond the US residential-material setting.

External indicators are aligned by quarter without strict reconstruction of their original release dates and historical data vintages. Consequently, chronological splitting alone does not establish performance under fully realistic information availability. RSMeans licensing also prevents public sharing of the underlying proprietary inputs.

The practical contribution is a structured way to forecast material-category input costs. Turning those forecasts into final project prices still requires overhead, profit, contingency and competitive-market considerations outside this study.