Online Tool Store
· 4 min read
How to Fix Estimates Without Estimating Better
Heshan Fernando
Co-founder & COO
You estimated three hours and it took seven. Next time you resolve to be more realistic, estimate five, and it takes eleven.
Trying harder does not work, and there is a good reason it does not. The bias is not a lack of care.
The planning fallacy is robust
People underestimate how long their own tasks will take. This holds when they know they usually underestimate, when they are reminded of specific past overruns, and when they are explicitly asked to account for the possibility of delay.
The finding is decades old and it replicates consistently. What makes it stubborn is that the estimate is generated by imagining the task going as planned, and imagining is the only tool available — you cannot imagine the specific unforeseen problem, because if you could foresee it you would have included it.
Interestingly, the bias largely disappears when estimating someone else’s task. Observers, lacking the inside view, estimate more accurately. That asymmetry is the clue: the problem is the method, not the effort.
Measure the bias instead of fighting it
If the error is systematic and in one direction, it can be corrected arithmetically.
Log estimate and actual for twenty or thirty tasks. Take the ratio of actual to estimate for each. The median of those ratios is your personal multiplier.
adjusted estimate = your estimate × your median ratio
For most people the ratio settles somewhere between 1.5 and 2.5, and it is remarkably stable — considerably more stable than any individual estimate. That stability is what makes the correction work.
You keep estimating exactly as before, badly, and multiply. The bad estimate is a useful input precisely because it is consistently bad.
| Approach | Works? |
|---|---|
| Try harder to estimate accurately | No — bias persists |
| Add a fixed buffer | Partly, if the buffer is right |
| Measure and multiply | Yes |
| Ask someone else to estimate | Surprisingly, yes |
Median, not mean
One task that overran by a factor of ten distorts a mean badly and tells you nothing about the typical case.
The median describes what usually happens, which is what you are applying to a typical future task. Keep the outlier in the log — it is informative about your risk profile — and use the median for the multiplier.
Separate ratios by task type
Once there is enough data, splitting by category is more useful than one number.
Familiar work estimates well. The ratio is often close to 1.2.
Novel work overruns badly, because there is no experience to estimate from and the unknowns are exactly what the estimate omits.
Work with dependencies on other people overruns worst of all, and the variance is enormous — the ratio is less useful than a range.
A single multiplier averages these, which over-inflates familiar work and under-inflates novel work. Two or three categories fixes that.
Estimate ranges, not points
A single number carries no information about confidence, and confidence varies enormously between tasks.
“Three days” for a familiar task and “three days” for something novel are very different claims presented identically. The first might be two to four; the second might be two to fifteen.
Giving a range communicates that. It also produces better planning decisions, because a task with a wide range is a task worth investigating before committing to a date.
The practical version is three numbers: best case, likely, and worst case. Where the worst case is several times the likely, that is a signal to spend an hour reducing the uncertainty rather than an hour arguing about the estimate.
Common mistakes to avoid
- Adjusting the estimate after the fact so the log looks better.
- Using the mean rather than the median.
- Applying one multiplier across very different kinds of work.
- Logging fewer than twenty tasks and trusting the ratio.
- Presenting the raw estimate externally while privately using the adjusted one.
How to do it with Estimation Calibrator
The Estimation Calibrator tracks the ratio over time.
- Log the estimate before starting and the actual when finished.
- Do not adjust either afterwards.
- After twenty tasks, read the median ratio.
- Multiply future estimates by it, and split by task type once you have the data.
Other productivity tools are in the tools directory.
Frequently asked questions
Why not just estimate better?
Because the planning fallacy persists even when people know about it and are reminded of past overruns. The estimate comes from imagining the task going as planned, and the delays are by definition unimagined.
Why the median rather than the mean?
Because one catastrophic overrun distorts a mean. The median describes your typical case, which is what applies to a typical future task.
Should the ratio differ by task type?
Yes, once you have enough data. Familiar work estimates far better than novel work or anything depending on other people, and one multiplier averages those into a figure that fits neither.
Final thought
Keep estimating badly and multiply. The bias is stable enough to correct arithmetically, which is more than can be said for your ability to overcome it.