A 2012 feasibility study commissioned by the Hong Kong Development Bureau established the first reference class forecasting (RCF) scheme in the Asia-Pacific region, building uplift curves for cost and duration from only 25 completed roadworks projects. The study's authors (Flyvbjerg, Hon and Fok, 2016) explicitly left several issues unresolved: the unexplained absence of Parkinson's-law consumption of contingencies in Hong Kong; the untested proposal to govern contingencies through a tiered, portfolio-level scheme; the integration of probabilistic thinking into incentive schemes and portfolio regimes; and the absence of guidelines for adjusting reference class forecasts without reintroducing bias. The study also records that its 25-project sample was a trade-off between the statistical validity of the results and the effort needed to collect the archival data, and its leave-one-out validation gave a 22% exceedance chance at P80 versus the 20% target, which the authors judged close enough to validate the class as robust. The subsequent RCF literature (through 2026, e.g. Bärenbold, 2023; Cantarelli et al., 2025) has not resolved these issues for Hong Kong, and the underlying methodological questions — how reference-class size, geographic scoping, weighting, adjustment and updating affect forecast reliability — remain open research problems in the field.

Background

Reference class forecasting (RCF) is a method for correcting optimism bias (the systematic tendency of early cost and schedule estimates to be too low) and strategic misrepresentation (deliberate underestimation to secure project approval) in infrastructure projects. Instead of building a bottom-up estimate, RCF takes what Kahneman and Tversky called the 'outside view': it collects a class of completed, similar projects, establishes the empirical probability distribution of their forecast errors (for example, actual cost divided by forecast cost), and reads off an 'uplift' — a percentage addition to the current estimate — corresponding to the decision maker's acceptable probability of overrun. A P80 uplift, for instance, is the uplift that historical experience suggests would be exceeded only 20% of the time. RCF has been adopted by the UK Department for Transport, endorsed by the American Planning Association, and applied in Denmark and elsewhere.

In 2012 the Hong Kong government's Development Bureau commissioned a feasibility study on RCF for major roadworks projects — the first such application in the Asia-Pacific region. Flyvbjerg, Hon and Fok (2016) reported its results. The reference class was built from 25 completed Hong Kong roadworks projects — a sample size the authors describe as a trade-off between the statistical validity of the results and the effort needed to collect the archival data — with forecast costs and durations compared against actual outcomes at successive stages of project development (the Hong Kong public-works 'Category' system, in which estimates mature from Category C, an early scheme-level estimate, through Category B to Category A, the estimate closest to construction start). Uplift curves for cost and duration were established at each category stage, and the 25 projects were benchmarked against an international database of 863 comparable highway projects. To test robustness, the authors used leave-one-out validation: each project in turn was withheld, a reference class was rebuilt from the remaining projects, and the withheld project's actual outcome was checked against the uplifts that class would have recommended. For Category C cost estimates this validation covered 18 projects: the P50 uplift would have been exceeded in 9 of 18 cases (consistent with the 50% target), while the P80 uplift would have sufficed in 14 of 18 — a 22% acceptable chance of overrun against the 20% target, which the authors read as close enough to validate the class as robust.

Two further observations motivate the unresolved issues. First, the study found no strong evidence of Parkinson's law — the adage, invoked in the paper, that resources given to managers will be fully consumed — in Hong Kong roadworks: half the sample underspent the budget estimated at the Category C upgrade, so contingencies were not used up, contradicting typical observations elsewhere and suggesting possible cultural or procedural differences that the authors state warrant further research. Second, unused contingencies imply an inefficient allocation of public funds (a pessimism bias in the full-cost benchmark of base cost plus contingency), so the authors proposed managing contingencies through portfolio management with a tiered scheme — delegated contract-level authority over roughly a P50–P55 contingency, project-level authority to about P60, and a portfolio-level pool covering total exposure up to P80 — but this scheme was proposed only, not empirically validated. The paper also notes that adjusting a reference class forecast (for example, claiming a project will perform better than the class average) risks reintroducing the very bias RCF was designed to remove, that integrating probabilistic thinking into project incentive schemes and portfolio regimes remained an untested enhancement, and that a standard data proforma was developed and project officers trained so that future projects' cost and time-to-completion data can be recorded and the uplift curves of the reference class updated 'at suitable intervals'.

The subsequent literature has not settled these questions. A systematic review of 41 RCF studies (Bärenbold, 2023) identifies the construction of the 'right' reference class — sufficiently similar projects, a class neither too small nor too large, with open questions about how geographic scope and project-similarity variables should be weighted — together with small-sample reliability, dynamic updating of classes, and hybridisation with other methods as the field's main open problems. Cantarelli, Davis, Pinto and Turner (2025) likewise state that questions about RCF's validity, applicability and limitations remain and set out a research agenda. Themsen (2019) contests RCF's accuracy from a longitudinal Danish megaproject case, and Zani and Adey (2025) validate a weighted variant of RCF on utility data while calling for further validation across industries and regions — evidence that class construction and validation methodology is active but unresolved. No published work updates or enlarges the Hong Kong roadworks reference class itself; the only later Hong Kong application found (Zhou, Zhang and Chiaradia, 2021) adapts RCF to appraise a large-scale urban planning vision, a different forecasting object.

Problem Statement

The 2016 Hong Kong study left several issues unresolved concerning its one-off feasibility-study reference class for public roadworks projects. The following research questions remain open:

  1. Explain or refute the absence of Parkinson's-law consumption of contingencies in Hong Kong roadworks — half the sampled projects underspent their Category C budget estimates, contradicting typical observations elsewhere — including the cultural or procedural factors behind it.
  2. Empirically validate (or refute) the proposed tiered, portfolio-level contingency scheme (delegated contract-, project- and portfolio-level contingencies covering cumulative risk exposure up to about P80) as a means of balancing budget exceedance against inefficient under-use of funds.
  3. Test the proposed integration of probabilistic thinking — the use of exceedance probabilities such as the P50/P80 uplift thresholds — into project incentive schemes and portfolio management regimes, in particular the tiered, portfolio-level contingency scheme the study proposes.
  4. Examine the reliability of the reference class and its uplift curves as the sample is enlarged beyond the original 25 projects — a size the authors describe as a trade-off between the statistical validity of the results and the effort needed to collect the archival data, and whose leave-one-out validation found the P80 uplift corresponded to a 22% acceptable chance of cost overrun against the 20% target, which the authors judged close enough to deem the class robust. The wider review literature (Bärenbold, 2023) additionally raises the question of how reference-class size and geographic scoping trade off against comparability of the projects in the class.
  5. Establish evidence-based guidelines for when and how reference class forecasts may be adjusted for project-specific information without reintroducing the optimism bias and strategic misrepresentation that the outside view is meant to remove.

These five issues jointly concern whether the one-off feasibility-study reference class can be developed into an empirically validated and maintainable forecasting-and-contingency-governance regime for Hong Kong roadworks.

References

  1. Flyvbjerg, B., Hon, C.-k., Fok, W. H. (2016). Reference class forecasting for Hong Kong's major roadworks projects. Proceedings of the Institution of Civil Engineers – Civil Engineering 169(6), 17–24. DOI: 10.1680/jcien.15.00075. Author version at https://arxiv.org/abs/1710.09419 (arXiv:1710.09419)
  2. Bärenbold, R. (2023). Reducing risks in megaprojects: The potential of reference class forecasting. Project Leadership and Society 4, 100103. DOI: 10.1016/j.plas.2023.100103. https://doi.org/10.1016/j.plas.2023.100103
  3. Cantarelli, C. C., Davis, K., Pinto, J. K., Turner, N. (2025). Reference class forecasting: promises, problems, and a research agenda moving forward. Production Planning & Control 37(7), 691–709. DOI: 10.1080/09537287.2025.2578708. https://doi.org/10.1080/09537287.2025.2578708
  4. Zani, D., Adey, B. T. (2025). Empirical comparison of weighted and non-weighted reference class forecasting. Infrastructure Asset Management 12(3), 147–158. DOI: 10.1680/jinam.24.00024. https://doi.org/10.1680/jinam.24.00024
  5. Themsen, T. N. (2019). The processes of public megaproject cost estimation: The inaccuracy of reference class forecasting. Financial Accountability & Management 35(4), 337–352. DOI: 10.1111/faam.12210. https://doi.org/10.1111/faam.12210
  6. Zhou, Y., Zhang, L., Chiaradia, A. J. F. (2021). An adaptation of reference class forecasting for the assessment of large-scale urban planning vision, a SEM-ANN approach to the case of Hong Kong Lantau tomorrow. Land Use Policy 109, 105701. DOI: 10.1016/j.landusepol.2021.105701. https://doi.org/10.1016/j.landusepol.2021.105701
  7. Chadee, A., Martin, H., Gallage, S., Rathnayake, U. (2023). Reducing cost overrun in public housing projects: A simplified reference class forecast for Small Island Developing States. Buildings 13(4), 998. DOI: 10.3390/buildings13040998. https://doi.org/10.3390/buildings13040998
  8. Sing, M. C. P., Kazeem, K. O., Chan, A. P. C., Liu, H. J. (2026). Synergies strategies for cost contingency: uniting reference class forecasting and probabilistic risk analysis in infrastructure project delivery. Production Planning & Control, 1–16. DOI: 10.1080/09537287.2026.2626435. https://doi.org/10.1080/09537287.2026.2626435