Explainable Machine Learning for Early-Stage Construction Duration Prediction Using Limited Project Information

Every construction project begins with a question that is difficult to answer without a crystal ball: how long will it take to build? Before the blueprints are fully drawn, before the materials are ordered, and before the first shovel hits the ground, contractors and clients must decide on a timeline. This decision dictates the budget, the schedule for workers, and the financial risk for everyone involved. Traditionally, answering this question required detailed engineering calculations and a breakdown of every single task, information that simply does not exist in the earliest planning stages. This creates a difficult situation where the need for a timeline is at its highest precisely when the data available to create one is at its lowest.

To solve this, researchers have turned to machine learning, a branch of computer science where programs learn to find patterns in data rather than following rigid, pre-written rules. In the context of construction, these programs can look at historical projects and learn how certain factors, like the cost or the type of building, relate to how long the work took. However, many previous attempts at this have relied on data that is only available after a project has already started, such as the exact number of floors or the specific site conditions. This limits their usefulness for the very early decisions that matter most. A new study by Mehmet Sena Kaşka and Işık Ateş Kıral addresses this gap by building a system that predicts construction duration using only the basic information available at the very start of a project.

The researchers gathered data from 209 completed construction projects to train their system. They deliberately chose to ignore complex details that are not yet known at the planning stage. Instead, they fed the computer four simple pieces of information: the estimated cost of the project, whether the owner is a public entity or a private company, the general type of building or infrastructure, and the geographic location. The dataset included a wide variety of structures, from residential homes and commercial offices to factories, ports, and water transmission lines, located across different regions including the Middle East, North Africa, South Asia, and the Commonwealth of Independent States. By using such a diverse set of projects, the team aimed to create a model that could handle the messy reality of the construction industry, rather than just a narrow slice of it.

To make sense of this data, the team tested several different types of machine learning algorithms. They compared models known as Random Forest, Extra Trees, Artificial Neural Networks, and XGBoost. These are different mathematical approaches to finding patterns, each with its own way of processing information. The researchers were careful not to rely on a single test to declare a winner. Instead, they repeatedly shuffled the data and tested the models in different ways to ensure the results were stable and not just a lucky accident of how the data was split. This rigorous testing revealed that while some models performed well on a single test, they were sensitive to changes in the data. The XGBoost model, which uses a technique of building many small decision trees and combining them, proved to be the most consistent performer across these repeated tests.

The final model selected by the researchers, XGBoost, achieved a level of accuracy that suggests it captured meaningful patterns in the data. When tested on a specific set of projects it had not seen before, the model's predictions were off by an average of about 175 days. While this might sound like a large margin of error, the projects in the study ranged in duration from as little as 92 days to as much as 2,760 days, with a median of 669 days. In the context of early planning, where no detailed schedule exists yet, this level of precision offers a valuable starting point. The model did not just provide a number; the researchers also used tools to explain how the model reached its conclusions, ensuring the results were not a "black box" mystery.

The analysis of the model's logic revealed that the estimated cost of a project was the single most important factor in predicting how long it would take. As the cost went up, the predicted duration generally increased, which aligns with the intuition that larger, more expensive projects tend to take longer. However, the model also learned that cost alone does not tell the whole story. The geographic location and whether the project was public or private also influenced the timeline. For instance, at similar cost levels, projects in certain regions were predicted to take longer than those in others. The type of project mattered as well, with infrastructure and port projects showing different duration patterns compared to factories or commercial buildings, though these factors were less influential than cost.

Despite these promising results, the authors are careful to define the limits of their work. They emphasize that this tool is not a replacement for a detailed construction schedule, which is created later in the process when all the specific plans are known. The model is designed as a preliminary aid, a way to benchmark a proposed timeline against historical data or to flag potential risks before a bid is submitted. The researchers also note that the model's performance is tied to the specific data it was trained on. Because the dataset contained only 209 projects and was unevenly distributed across different types and locations, the model might behave differently if applied to a completely new country or a different market condition. The study explicitly states that the model needs to be tested on new, external data before it can be trusted for critical real-world decisions.

Ultimately, this research demonstrates that even with very limited information, it is possible to make informed guesses about construction timelines. By focusing on the few variables that are known at the very beginning of a project, the team created a framework that helps decision-makers navigate the uncertainty of the early planning phase. The study shows that while a computer cannot predict the future with perfect certainty, it can learn from the past to provide a reasonable estimate that is better than a guess. This approach offers a practical way to bring data-driven insights to the earliest stages of construction, helping to manage expectations and plan resources before the first brick is laid.