4 ms·

A lot of this stems from the accepted norm that an estimate is a single value. This loses a lot of information and when combined with other lossy estimates, thi

A lot of this stems from the accepted norm that an estimate is a single value. This loses a lot of information and when combined with other lossy estimates, this error compounds. Pair any estimate (time, cost, effort) with a confidence indicator. This will likely improve as more about the problem is known, but until then gives the ability to reason with appropriate certainty based on a set of inherently noisy priors. There’s a whole set of more formal systems such as PERT (https://en.m.wikipedia.org/wiki/Program_evaluation_and_review_technique https://en.m.wikipedia.org/wiki/Program_evaluation_and_revie...) which base off this principle. That doesn’t work for every project, but there’s often an over correction in ‘modern’ software teams where significant research, experience, and process that has been designed to avoid project chaos is completely ignored for the sake of being ‘lean’.

valenterry 4y ago
Exactly. As a minimum, it should be something like "90% chance it will be between 3 and 10 weeks, 50% chance it will be between 5 and 6 weeks". Optimally we have a probability distribution. Not to forget that over time the confidence will grow and estimates become more precise.
kqr 4y ago
A single value isn't the problem -- the lack of an indication of confidence is. If I say "there's a 90 % chance it's done before March 13." that's a useful and verifiable estimation, even though it's only one number.
- brazzy 4y ago
  It's two numbers.
osigurdson 4y ago
>> A lot of this stems from the accepted norm that an estimate is a single value. If we want to estimate accurately, it is necessary to break work down into smaller pieces. However, we can only estimate the items that we can identify at the time of estimation (I.e., we know that we need to do A,B and C but discover that D,E and F are needed as well). The other factor is combinatorial explosion: we'll do A,B or C and this impacts what we do next. This graph can become enormous when attempting to construct up-front. I agree that having a probability distribution for each estimate is better than a single value. If you have two values, this is a uniform distribution. It is also possible to do a normal distribution, triangular, truncated normal, etc. However, it isn't clear to me that increased sophistication leads to better outcomes.
- kimburgess 4y ago
  Completely agree. The overcorrection applies in both directions. The two value (estimate and confidence) is just a simple approach that is rather intuitive to interpret directly. As you note, diving into 3 point estimates gives to ability to extract beta distribution or other weightings that may or may not be suitable for the given domain. Complete top down planning for any work that involves discovery is useless, at best. That doesn't mean no planning is better. (Source: have experienced and been actively involved in both sides of that overcorrection.)

valenterry 4y ago

Exactly. As a minimum, it should be something like "90% chance it will be between 3 and 10 weeks, 50% chance it will be between 5 and 6 weeks". Optimally we have a probability distribution. Not to forget that over time the confidence will grow and estimates become more precise.

kqr 4y ago

A single value isn't the problem -- the lack of an indication of confidence is. If I say "there's a 90 % chance it's done before March 13." that's a useful and verifiable estimation, even though it's only one number.

brazzy 4y ago
It's two numbers.

osigurdson 4y ago

>> A lot of this stems from the accepted norm that an estimate is a single value. If we want to estimate accurately, it is necessary to break work down into smaller pieces. However, we can only estimate the items that we can identify at the time of estimation (I.e., we know that we need to do A,B and C but discover that D,E and F are needed as well). The other factor is combinatorial explosion: we'll do A,B or C and this impacts what we do next. This graph can become enormous when attempting to construct up-front. I agree that having a probability distribution for each estimate is better than a single value. If you have two values, this is a uniform distribution. It is also possible to do a normal distribution, triangular, truncated normal, etc. However, it isn't clear to me that increased sophistication leads to better outcomes.

kimburgess 4y ago
Completely agree. The overcorrection applies in both directions. The two value (estimate and confidence) is just a simple approach that is rather intuitive to interpret directly. As you note, diving into 3 point estimates gives to ability to extract beta distribution or other weightings that may or may not be suitable for the given domain. Complete top down planning for any work that involves discovery is useless, at best. That doesn't mean no planning is better. (Source: have experienced and been actively involved in both sides of that overcorrection.)