Yield Curve Forecasting

Definition

Yield Curve Forecasting in project controls is the practice of extrapolating remaining cost, duration, and physical output by fitting an empirical yield or learning curve to the project's performance to date. Common forms include the Gompertz, logistic, and Weibull S-curves for cumulative cost and progress, and the Wright-Crawford log-linear curve for unit-rate productivity improvement over repetitive work. Underlying terminology is maintained in the PMMilestone PM Glossary.

History

The learning-curve effect was first quantified by T. P. Wright in 1936 for aircraft assembly and refined by J. R. Crawford in 1944. Cumulative S-curve forecasting was formalised by the U.S. Army Corps of Engineers in the 1960s and is referenced in AACE International Recommended Practice 17R-97 (Cost Estimate Classification System). The Construction Industry Institute's productivity research, ongoing since the 1990s, supplied much of the empirical data behind modern yield-curve practice. The historical development is taught in depth in the forecasting module of the Project Controls Academy.

Principles

• Fit the curve to actual data, not to the baseline — the baseline is the plan, not the forecast.
• Use at least 20–30 percent progress before trusting the extrapolation; early data is too noisy.
• Document the curve type, fit statistic (R²), and outlier handling — a forecast without a residual diagnostic is a guess.
• Reconcile the curve forecast with the bottom-up Estimate at Completion (EAC) from the EVM Calculator; large divergence is itself a signal.
• Refresh every reporting period and track the trend in the fitted parameters, not just the latest forecast point.

Real-World Construction Example

On a 14 km tunnel boring project, the contractor's baseline assumed 22 m/day advance after a six-month ramp-up. By month nine the actual was 17 m/day with an upward trend. The controls team fitted a Gompertz S-curve to monthly cumulative metres bored. The fitted asymptote forecast a completion 11 weeks later than the baseline. The owner used the forecast to renegotiate downstream contracts before they became commercial issues. Three months later the actual completion projection had closed to within two weeks of the Gompertz fit. The discipline of trusting the data over the plan saved a costly schedule recovery exercise.

IT / Agile Example

A software platform team had a 480-story-point backlog and three months of velocity data. Naive linear extrapolation suggested a 14-sprint completion. Fitting a Crawford log-linear learning curve — accounting for productivity improvement as the team mastered the domain — projected 11 sprints. The team published both forecasts and tracked which one tracked actuals. By sprint 5 the Crawford forecast was within 6 percent; the linear was off by 22 percent. The portfolio team adopted Crawford fits for all new mature-domain teams.

Project Controls Perspective

Yield-curve forecasting is the analytical complement to EVM. Where EVM applies indices to remaining work, yield curves let the data describe its own future shape. The two should agree within a narrow band; persistent divergence flags a data-quality issue, a method bias, or an undisclosed scope change. Validate forecasts against the CPI Calculator, the SPI Calculator, and the Schedule Health Checker.

Practical Lessons Learned

• Curve fitting is not magic. Garbage data produces a confidently wrong line.
• Always show the confidence band, not just the central forecast. Reporting only the point estimate is dishonest.
• If the curve refuses to fit, the process is non-stationary — investigate before forcing a fit.
• Owners listen to forecasts that have been right before. Track your own forecast accuracy publicly.

Common Mistakes

• Extrapolating from a non-stationary process — for example after a major scope change or crew turnover.
• Reporting only the central forecast and hiding the confidence band.
• Choosing the curve shape that gives the answer you want, rather than the one with the best fit statistic.
• Refitting only at milestones rather than every reporting period.

Expert Tips

• Use a symmetric logistic as the default S-curve, Gompertz for slow-starting tunnelling and modular fabrication, and Weibull when the tail is long.
• Plot the fitted curve alongside the baseline curve and the actuals. The visual carries the message faster than tables.
• Audit your own forecasts six months later. Adjust the model — not the truth — when it is consistently off.

Key Takeaways

• Yield-curve forecasting lets the data project its own completion.
• Used alongside EVM EAC and bottom-up ETC, it triangulates the most defensible forecast.
• Confidence bands and fit statistics are mandatory honesty checks.
• The discipline rewards consistency more than cleverness — fit, refresh, report, audit.

Further Reading

Practitioner references on S-curve and learning-curve forecasting are listed in PMMilestone Books & Publications.