State of the Practice in Software Effort Estimation: A Survey and Literature Review

Effort estimation is a key factor for software project success, defined as delivering software of agreed quality and functionality within schedule and budget. Traditionally, effort estimation has been used for planning and tracking project resources. Effort estimation methods founded on those goals typically focus on providing exact estimates and usually do not support objectives that have recently become important within the software industry, such as systematic and reliable analysis of causal effort dependencies. This article presents the results of a study of software effort estimation from an industrial perspective. The study surveys industrial objectives, the abilities of software organizations to apply certain estimation methods, and actually applied practices of software effort estimation. Finally, requirements for effort estimation methods identified in the survey are compared against existing estimation methods.

💡 Research Summary

The paper presents a comprehensive industrial‑focused investigation of software effort estimation, a discipline that remains pivotal for delivering projects on time, within budget, and at the agreed quality level. While traditional estimation techniques have been designed primarily for planning and tracking and therefore emphasize point‑value accuracy, contemporary software organizations increasingly demand capabilities that go beyond a single number: systematic analysis of causal effort drivers, explicit representation of uncertainty, and continuous learning from past projects.

To capture the state‑of‑practice, the authors conducted a large‑scale survey targeting a diverse set of companies (large enterprises, mid‑size firms, and small vendors) across multiple domains and development methodologies. The questionnaire was structured around four themes: (1) estimation objectives (accuracy, causal analysis, reliability, trade‑off balancing among cost, schedule, and quality); (2) organizational capabilities (availability and quality of historical project data, data‑management processes, expertise of estimation staff, tooling support); (3) currently employed estimation methods (expert judgment, COCOMO, Function Point Analysis, statistical regression, machine‑learning models, hybrid approaches); and (4) satisfaction levels and perceived gaps. Over 150 organizations responded, providing a rich data set that reflects real‑world constraints and aspirations.

Key findings from the survey are as follows. First, while “accurate point estimates” remain the top‑ranked goal, there is a pronounced shift toward “causal‑relationship analysis” and “quantified confidence” – 68 % of respondents indicated that understanding why effort changes and being able to express uncertainty are now critical for risk management and stakeholder communication. Second, the ability to leverage historical data varies dramatically: large firms often have mature data warehouses and automated pipelines, whereas many SMEs lack systematic data collection, leading them to rely heavily on expert judgment. Third, the distribution of methods in practice shows a dominance of traditional techniques – expert judgment (≈45 % of projects), COCOMO (≈30 %), and Function Point Analysis (≈20 %). Machine‑learning‑based approaches are used in less than 5 % of cases, primarily due to concerns about data quality, model interpretability, and integration effort.

From these observations the authors distilled four minimum requirements that any modern effort‑estimation method should satisfy: (1) Causal Modeling – the method must explicitly capture relationships between effort and drivers such as requirement volatility, staff turnover, technical debt, and tool usage; (2) Uncertainty Quantification – estimates should be accompanied by confidence intervals, probability distributions, or Bayesian posterior estimates; (3) Continuous Learning – the technique must be able to ingest new project data and update its parameters without extensive re‑engineering; (4) Methodology‑Tool Agnosticism – the approach should integrate seamlessly with waterfall, agile, DevOps, and hybrid processes, preferably via open APIs or plug‑ins for popular ALM tools (Jira, Azure DevOps, etc.).

The paper then conducts a systematic literature review, mapping a representative set of established estimation techniques against the four requirements. Traditional parametric models such as COCOMO and Function Point Analysis excel at providing quick, calibrated point estimates but fall short on causal transparency and uncertainty reporting. Regression‑based statistical models can incorporate multiple drivers but are highly sensitive to data quality and typically produce only confidence bands, not a full causal graph. Recent machine‑learning models (e.g., random forests, neural networks, Bayesian networks) are capable of modeling complex, non‑linear interactions and can generate probabilistic outputs, yet they suffer from a lack of interpretability, high data‑volume demands, and limited out‑of‑the‑box integration with project‑management ecosystems.

Given that no single existing technique satisfies all four industrial requirements, the authors propose a Hybrid Estimation Framework. The framework operates in three layers: (a) a baseline layer that uses expert judgment and calibrated parametric models to produce an initial estimate quickly; (b) a data‑driven refinement layer that applies statistical or machine‑learning models to historical project data, updating the baseline and providing probabilistic adjustments; (c) an uncertainty & causal visualization layer that employs Bayesian inference or causal‑graph techniques to surface driver‑effect relationships and generate confidence intervals that can be directly embedded into project dashboards. This architecture leverages the strengths of human expertise (contextual insight, rapid boot‑strapping) while exploiting the objectivity and learning capacity of data‑driven models.

The authors discuss practical implications for software organizations. First, establishing a robust data‑collection and cleansing pipeline is a prerequisite; without reliable metrics on size, effort, defect density, and process attributes, even the most sophisticated models will produce misleading results. Second, cross‑functional teams comprising estimation experts, data scientists, and project managers should be formed to maintain the hybrid system, ensuring that model updates reflect both statistical evidence and domain knowledge. Third, tool support must be prioritized: the estimation engine should expose RESTful services or plug‑ins that can be called from existing ALM platforms, enabling real‑time “what‑if” analyses during sprint planning or release forecasting. Fourth, continuous training programs are essential to keep staff familiar with both traditional estimation concepts and emerging analytics techniques, reducing resistance to adoption.

In conclusion, the study provides an evidence‑based snapshot of current industrial practices, highlights a growing mismatch between legacy estimation methods and emerging business needs, and offers a concrete roadmap—a hybrid, uncertainty‑aware, causally transparent framework—to bridge that gap. The authors call for further empirical research to validate the proposed framework in longitudinal industrial settings and to explore automated extraction of causal relationships from version‑control and issue‑tracking data.