Your finished projects
Use one unit for both columns, hours or dollars. Leave out projects still in progress. Rows are read in the order given (oldest first) so the back-test can score your most recent projects. Needs at least 12 finished projects; 25 or more for a usable answer. An optional third column (type of work) adds an overrun per type once a type has 5 projects.
Upload or paste your projects, then click Analyse.
Whole-project total from your task list
Result
How your estimates have missed
What to quote for this estimate
Each row: the amount that share of comparable past projects came in at or under. The loss chance columns need your rate.
| Covers | Expected actual | Buffer over estimate | Price to cover it | Chance a price at that level still loses |
|---|
By type of work
Back-test
How it is calculated
Model: ln(actual) = a + b × ln(estimate) + noise. By default b = 1, so each project is an overrun ratio and a is the median log ratio; b is fitted only when there are 20 or more projects and the slope differs from 1 by more than two standard errors (overruns that change with size). The 50% figure is the median. The 70 to 95% figures add the ceil((n+1)×p)-th smallest residual, the finite-sample correction used in split-conformal prediction, so no bell curve is assumed, and a level the sample cannot support (for example 95% with 15 projects) is flagged. Fixed-price loss chance: the share of residuals that would push actual × rate above the price, with a +1 so a short history never claims zero risk. The 80% ranges come from 300 seeded bootstrap resamples of your projects, so a result is reproducible. Back-test: the first 75% of rows fit the model, the last 25% are scored on how many land at or under the predicted 80% level (should be near 80%) and whether the model beats assuming actual equals estimate.
Not modelled: scope changes agreed with the client, a different kind of work than you have done before, changes in how you estimate, correlation between projects for the same client, and unbilled time outside the actual you entered. Unverified: checked against simulated histories with known overruns in unit tests, not against a published dataset or a commercial tool.
Common questions
How much should I pad my project estimates?
It depends on how your own estimates have missed. This tool takes your finished projects, measures actual divided by estimate, and reports the multiple that covered 50%, 70%, 80% and 90% of them. Quoting at the 80% level means roughly one project in five still runs over; the tool shows the buffer that level implies for your history instead of a generic 20% or 30% rule.
Why do my projects always run over the estimate?
Usually because estimates describe the path where nothing goes wrong, while actual time includes rework, waiting for feedback and scope that was not written down. The miss is lopsided: a project can take twice as long but never half as long. The tool shows how often you ran over, your typical overrun and your worst case, so the pattern is measured rather than remembered.
What is a P80 estimate?
It is the amount that 80% of comparable past projects came in at or under. A P50 is the coin-flip number, so a fixed price set at your P50 loses money about half the time. A P80 price still loses on roughly one project in five, but the losses are far rarer, and the buffer is paid for by the ones that finish on time.
How many finished projects do I need?
At least 12, and 25 or more for a result the tool will call usable. With fewer than about 20 projects the 95% level cannot be estimated beyond your worst one, so the tool says so. More history, especially projects of similar kinds, gives a tighter range.
Does it work for dollars as well as hours?
Yes. Use one unit for both columns: estimated and actual hours, or estimated and actual cost. The overrun ratio is what matters. To see the chance a fixed price loses money you also enter your hourly cost or rate, so the tool can turn hours into money.
Why might the result be wrong?
Only your past shows up here. If the next project is a different kind of work, a new client or a new technology, the history may not apply, and the tool flags estimates outside the range you have done before. Your estimating habits may also have changed, which the hold-out back-test on your most recent projects helps to reveal. Treat the answer as a starting buffer, not a guarantee.
Is my project history uploaded anywhere?
No. The file is read and analysed in this browser tab. Nothing is sent to a server, there are no accounts or cookies, and the tool never stores your projects. It works offline once the page has loaded.