Reference class forecasting

The outside view: building a library of comparable projects and testing your contingency against it.

Note

Reference class forecasting is in early access and is switched on for an organisation by us. If the RCF (Outside view) page is missing from your sidebar, ask for it to be enabled.

What it is for

The cost and schedule analyses build up from this project’s own risks. That is the inside view, and it is only as good as the risks someone thought to write down. Reference class forecasting is the outside view: it ignores the detail of this project and asks how comparable completed projects actually finished against their estimates, at the same stage of maturity. Projects tend to overrun more than their registers suggest, and this is how that shows up.

It answers a different question from the simulation. The risk models say why a project could overrun; the outside view says how far projects like it did. Using both is stronger than either.

The reference library

The library is shared by every project in your organisation and lives on the Reference library tab. It is built from reference classes: groups of work whose past projects are a fair comparison for each other, such as substations or underground cable. Name a class for the type of work, not for a client.

Each class can draw on three kinds of evidence, which are always kept apart:

Company
Your own completed projects. For each, enter what it was estimated at, at one or more business-case stages (strategic outline, outline, full), and what it finally cost and took. You can paste a table to load many at once. These count as soon as they are added. A class needs at least five projects at a stage before it produces a curve.
Licensed
Percentile curves your organisation is licensed to use from a third party. You paste them in with a note of who they are licensed from. They stay private to your organisation and are never sent to an AI model.
Global (AI research)
Completed public projects found by AI research. See the next section.

Who can do this

Changing the library needs the organisation reference-data permission, which Owners and Admins have. Anyone who can view the project can see the results.

AI research

Inside a class, Research comparable projects searches public sources for completed projects of that kind that state both an estimate and an outturn. You can steer the search, for example by region or size. It can take a few minutes.

Nothing it finds is trusted. Each finding links to its source and is held as to review. It counts towards a forecast only after a person has opened the source and accepted it. Rejected findings are remembered so they do not keep coming back. Public reports over-represent well-known overruns, so global figures are best used as a cross-check rather than as the basis for a funding decision.

Assessing a project

On the Assessment tab, work down the page:

  1. Choose the evidence to use. Company, Licensed and Global are shown side by side and never merged into one figure.
  2. Score the project’s maturity. Ten criteria, from business case to team, each scored from not started to in delivery. The average decides which stage of evidence applies, and results move smoothly between stages rather than jumping. Score honestly: an optimistic score puts you on a more favourable curve and defeats the point.
  3. Enter the cost to go. Split the remaining cost by type of work and tag each line with a reference class. Count future cost only: leave out money already spent, and leave out the contingency itself.
  4. Enter the contingency held, so the tool can tell you how confident you can be in it.
  5. Enter the schedule to go. Add key activities with planned dates and tag them with a class. Work finished before the analysis date is ignored. Enter the float held in days.

Reading the results

Switch between Cost and Schedule. For each evidence source you see:

  • P50 and P80 uplift: how far comparable projects overran their estimate, as a percentage and as a figure on your cost to go or remaining duration.
  • Confidence in the current contingency or float: the share of comparable projects that would have stayed within what you hold.
  • Evidence coverage: the share of your cost or remaining work whose class has evidence in that source.

Notices flag thin evidence (fewer than five projects), work with no class, and a project that is less or more mature than the evidence available. Take them seriously: they say how far the tails can be trusted. How classes are combined lets you assume every class overruns together, which gives the widest range, or a lower setting if your classes are genuinely separate kinds of work.

Snapshots and history

Save snapshot records the inputs and the results from every source as they stand today. Snapshots are never edited. The History tab shows how the outside view has moved as the project matured, and the latest snapshot is what the reports and the QCRA chart use.

Alongside the risk models

Once a snapshot is saved, the QCRA confidence curve gains a second line showing what comparable past projects needed, with a plain statement of the gap at the top confidence level. A wide gap means the register may be missing risks or carrying optimism.

Reports

Two reports on the Reports page are built from the latest snapshot:

  • Outside view (RCF): the contingency and float at each confidence level, the results by evidence source, cost to go by class, and how the forecast has moved across snapshots.
  • Inside vs outside: the risk register and schedule risk model set beside comparable past projects, with the gap at P80 and a plain-English reading.

Both export to PDF. See Reports for the general workflow.

Still stuck?

Tell us what you were trying to do and what happened instead. A screenshot of the page helps more than a description of it.