💻 computer science

Predictive Intelligence for Civil Works Valuation: A Hybrid Random Forest–ANN Model with 91.20% Accuracy for Dynamic Cost Management in Peru

This study proposes a hybrid Random Forest–ANN model trained on 450 Peruvian construction projects that achieves 91.20% accuracy in dynamic cost valuation, significantly reducing prediction errors and geographic bias while offering a viable pathway for integration into Peru's public infrastructure management frameworks.

Original authors: PAUL RICARDO PRUDENCIO GALVEZ

Published 2026-06-30

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: PAUL RICARDO PRUDENCIO GALVEZ

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to guess the final price tag of building a house in Peru. In the past, people did this by looking at a static list of prices (like a menu from last year) and adding them up. But in Peru, the "menu" changes constantly because of three tricky things: the weather and terrain (coast vs. mountains vs. jungle), the altitude (how high up you are), and the risk of earthquakes.

Because of these changing factors, traditional guessing methods were often wrong by about $142,800 per project. That's a huge financial gap.

This paper introduces a new, smarter way to predict costs using a "digital brain" that learns from the past. Here is the breakdown in simple terms:

1. The Problem: The "One-Size-Fits-All" Map Doesn't Work

Imagine trying to navigate Peru using a single map that treats the beach, the high Andes mountains, and the Amazon jungle exactly the same. It wouldn't work. A truck driving up a mountain costs more fuel than one on the coast; a building in an earthquake zone needs stronger (and more expensive) materials.

The old computer models used in construction didn't understand these differences. They treated the whole country as one flat, uniform place, leading to expensive mistakes.

2. The Solution: A "Double-Brain" System

The author created a new computer model that acts like a two-person expert team working together:

Brain A (Random Forest): This part is great at looking at many different rules and patterns at once. It's like a veteran engineer who has seen thousands of projects and knows the "rules of thumb."
Brain B (Artificial Neural Network): This part is like a pattern-recognition wizard. It looks at complex, messy data and finds hidden connections that humans might miss.

By combining these two brains, the model gets the best of both worlds. It was trained on 450 real construction projects in Peru from 2018 to 2025.

3. The Secret Sauce: The "Peruvian Adjustment Factors"

The model doesn't just look at the price of bricks and cement. It adds a special "Peruvian filter" to every calculation. It asks three specific questions before giving a price:

Where is it? (Coast, Sierra, or Jungle?)
How high is it? (Is it above 3,500 meters, where logistics are harder?)
Is it shaky? (What is the earthquake risk level according to Peru's building codes?)

This is like a GPS that doesn't just tell you the distance, but also calculates how much extra gas you need because of the steep hills and traffic jams.

4. The Results: A Much Sharper Crystal Ball

The results were impressive:

Accuracy: The new model predicted costs with 91.20% accuracy.
Money Saved: Instead of being off by $142,800 (the old way), the new model is only off by about $91,350. That's a saving of roughly $51,450 per project.
Bias Reduced: It cut the "geographic bias" (the error caused by ignoring location differences) by 34%.

5. Why This Matters for Peru

The paper argues that this isn't just a math trick; it's a practical tool for the government and private companies.

Real Data: It uses real public records from Peru's contracting system (SEACE).
Future-Proof: It fits into Peru's existing digital systems (like PlanBIM) and follows the country's specific laws on how construction prices should be adjusted.
Less Drama: By predicting costs more accurately, there will be fewer arguments and renegotiations when a project goes over budget.

In short: The author built a smart computer tool that understands the unique, rugged, and earthquake-prone reality of Peru. By combining two types of AI and adding specific local rules, it predicts construction costs much more accurately than the old methods, saving the country significant money and reducing financial risk.

Technical Summary: Predictive Intelligence for Civil Works Valuation in Peru

Problem Statement
Civil engineering projects in Peru face significant challenges in dynamic cost valuation due to input price volatility, regulatory complexity under the State Contracting Law (Law N° 30225), and the inherent subjectivity of traditional parametric estimation. While global machine learning (ML) models exist, they fail to address the specific normative, geographic, and logistical realities of the Peruvian market. These realities include extreme geographic dispersion (coast, sierra, jungle), distinct seismic zones defined by NTE E.030, altitude variations, and a readjustment structure based on INEI polynomial formulas. Consequently, traditional methods yield average prediction errors exceeding ±USD 142,800 for medium-scale projects, with infrastructure projects experiencing cost overruns exceeding 58% and delays of approximately 65%. Existing approaches, often relying on univariate LSTM or linear regression, cannot simultaneously capture the relational and temporal dependencies required to model the cascade effects of activity deviations within a Work Breakdown Structure (WBS).

Methodology
The study proposes a hybrid ML model integrating Random Forest (RF) and Artificial Neural Networks (ANN) to address these gaps. The methodology is grounded in a dynamic cost formulation ( $C_d$ ) that surpasses static algebraic summation:

Data Source: The model was trained on 450 construction projects executed in Peru between 2018 and 2025, sourced from SEACE, the Ministry of Economy and Finance (MEF), and regional public entity records. Input variables included built area, structural type, geographic location, logistical complexity, seismic zone, altitude, local material price history, and regional CPI series.
Dynamic Cost Equation: The core equation is defined as $C_d(t) = \Sigma [Q_i(t) \times P_{base,i} \times I_{regional}(t)] + \epsilon_{ML}$ $C_{d} (t) = Σ [Q_{i} (t) \times P_{ba se, i} \times I_{r e g i o na l} (t)] + ϵ_{M L}$ .
- $Q_i(t)$ : Physical quantities extracted from 5D BIM models.
- $P_{base,i}$ : Homologated base unit prices from CAPECO and MEF.
- $I_{regional}(t)$ : A spatio-temporal regional index calculated as $I_{nacional}(t) \times \phi_{geo} \times \phi_{alt} \times \phi_{sísmica}$ . This index uniquely incorporates geographic factors (coast/sierra/jungle), altitude corrections (>3,500 m.a.s.l.), and seismic zone adjustments (NTE E.030).
- $\epsilon_{ML}$ : A predictive correction factor generated by the hybrid model to capture residual errors not explained by official indices.
Hybrid Architecture: The model utilizes a weighted hybrid loss function ( $L_{hybrid} = \alpha \cdot MSERF + (1-\alpha) \cdot MAEANN + \lambda||\theta||^2$ ) with $\alpha=0.6$ to balance variance control (RF) and bias reduction (ANN). L2 regularization ( $\lambda$ ) prevents overfitting given the limited dataset size ( $n < 600$ ).
Validation: Statistical stability was confirmed via 10-fold cross-validation ( $k=10$ ), yielding a standard deviation of $\sigma = \pm 2.1\%$ .

Key Contributions
The paper identifies three primary contributions:

Algorithmic Development: The creation and validation of a hybrid RF+ANN algorithm specifically trained on Peruvian data, achieving 91.20% accuracy.
Mathematical Formalization: The formulation of mathematically verifiable dynamic cost equations, including the spatio-temporal regional index ( $\phi_{geo} \cdot \phi_{alt} \cdot \phi_{sísmica}$ ) and the integration of Earned Value Management (EVM) indicators (CPI, EAC) as feedback variables.
Implementation Framework: A proposed pathway for integrating this model with the PlanBIM Peru ecosystem, SEACE, and the MEF Multiannual Macroeconomic Framework to support EPC contracts and Public-Private Partnerships (PPPs).

Results
The hybrid model achieved 91.20% accuracy on the test set (90 projects), with an $R^2$ of 0.9819 and an RMSE of 0.1358.

Error Reduction: The average prediction error was reduced from ±USD 142,800 (traditional methods) to ±USD 91,350 (ML model), representing a 35.9% reduction in estimation financial risk and an absolute improvement of USD 51,450 per project.
Bias Mitigation: The integration of the spatio-temporal regional index reduced geographic bias by 34% compared to models using a single national index (statistically significant, $p < 0.001$ ).
Performance Context: The model outperformed traditional parametric estimation by 18.8 percentage points and demonstrated superior performance in high-inflation contexts (2022–2023) and high-altitude sierra projects.

Significance and Claims
The paper claims that this study sets a new standard for financial management in Peruvian infrastructure by bridging the gap between global ML capabilities and local regulatory/geographic specificities. The 91.20% accuracy is presented not merely as a statistical metric but as a tangible mitigation of financial risks in EPC and PPP contracts, where cost overruns often stem from inadequate regional readjustment forecasting.

The authors assert that the model is operationally viable within the current Peruvian framework, leveraging existing public data transparency (SEACE) and government initiatives (PlanBIM, Invierte.pe). However, the paper modestly acknowledges limitations, including the heterogeneity of historical data formats from regional entities, the absence of a national georeferenced unit cost repository, and the necessity for semi-annual model updates to accommodate regulatory changes in the National Building Regulations. Future research is directed toward data standardization and extending the model to schedule prediction (Earned Schedule).