Saeed AzarshabResearch & publications

P80: The Making of a Forecast

4 min 27 sec · English narration and captions · Published

Watch how Mohammadsaeed Azarshab used a construction simulation to test whether P80 cost and finish-date forecasts stay calibrated when activities share risk. The 3D scenes are visual reconstructions of the research process.

Follow the study from the planning question through simulated project networks, the test of three forecasting approaches and the limits of the results. The manuscript was submitted to a journal on 15 November 2026.

Read about the P80 construction cost and schedule study

Read the English transcript
  1. A project manager puts one date and one budget in front of a client. How much confidence should either number earn?
  2. Mohammadsaeed Azarshab turned that planning question into a test of the forecasts themselves.
  3. Picture a hundred similar projects. A well calibrated P80 budget should cover about eighty of them when outcomes vary continuously.
  4. Now imagine one late delivery holding up several trades. Their risks move together. Does the forecast allow for that?
  5. Azarshab reviewed earlier work on project risk and forecast accuracy, then narrowed the question to a comparison he could repeat.
  6. Synthetic networks let him control exactly which activities shared risk, and check each forecast against the simulated outcome.
  7. The core design used sixty activities in six groups. The study varied how the tasks were connected and how strongly their risks moved together.
  8. At six checkpoints, the forecaster learned more about the project and updated its budget and finish date.
  9. The first method treated risks in different activities as independent. The second knew the dependence built into the simulation.
  10. The third compared six candidate models and kept the highest P80 forecast. A higher allowance can also carry a cost.
  11. Before the main run, Azarshab registered four hypotheses, their thresholds, and the rules for analysing cost and duration.
  12. That written record mattered later. It specified exactly what evidence would count if the two outcomes told different stories.
  13. The registration exposed problems with text encoding, the attachment path and the platform response. Each correction received its own record.
  14. When the archived package could not be retrieved as expected, an alternative repository preserved a verifiable copy.
  15. A second obstacle came from the certificate used to verify a public randomness pulse. The main run stopped at that check.
  16. After the pulse was known, a dated amendment documented the authentication change before computation. The designated pulses and seed rules were retained.
  17. After development tests and release checks, the engine ran in blocks. It could pause, check completed work, and resume.
  18. The completed run covered 14,400 core trajectories and 4,000 more for sensitivity checks.
  19. The audit checked 331,200 forecast rows. At each state, it checked the envelope against all six candidates.
  20. 10,000 bootstrap resamples kept each network together when estimating uncertainty.
  21. With strong shared risk, the independence model covered final cost in about sixty five per cent of simulated cases.
  22. It covered duration in about seventy two per cent. Both fell below the intended eighty per cent.
  23. The envelope reduced the average calibration gap for cost by about four percentage points. For duration the gap grew by nearly two.
  24. It improved the forecast loss score for both outcomes. But the fourth hypothesis required every condition to hold. It was not supported.
  25. A separate diagnostic began with a timing pilot and expanded to 48,000 artificial datasets.
  26. Two-sided intervals showed undercoverage in 23 of 24 simplified settings. That limits what the intervals can establish.
  27. The findings became tables, figures and a manuscript. Azarshab reviewed the source claims and kept the mixed fourth result in the paper.
  28. He also prepared an anonymous data and code archive so a reviewer could check the numerical work.
  29. On the fifteenth of November, twenty twenty six, the International Journal of Forecasting received the manuscript and its supporting files.
  30. For a planner, the question to ask is simple. Which activities could hit the same obstacle, and does our forecast account for them?