17 Theses on Software Estimation

(with apologies to Martin Luther for the title)

Arriving late to the #NoEstimates discussion, I’m amazed at some of the assumptions that have gone unchallenged, and I’m also amazed at the absence of some fundamental points that no one seems to have made so far. The point of this article is to state unambiguously what I see as the arguments in favor of estimation in software and put #NoEstimates in context.

1. Estimation is often done badly and ineffectively and in an overly time-consuming way. 

My company and I have taught upwards of 10,000 software professionals better estimation practices, and believe me, we have seen every imaginable horror story of estimation done poorly. There is no question that “estimation is often done badly” is a true observation of the state of the practice.

2. The root cause of poor estimation is usually lack of estimation skills. 

Estimation done poorly is most often due to lack of estimation skills. Smart people using common sense is not sufficient to estimate software projects. Reading two page blog articles on the internet is not going to teach anyone how to estimate very well. Good estimation is not that hard, once you’ve developed the skill, but it isn’t intuitive or obvious, and it requires focused self-education or training.

3. Many comments in support of #NoEstimates demonstrate a lack of basic software estimation knowledge. 

I don’t expect most #NoEstimates advocates to agree with this thesis, but as someone who does know a lot about estimation I think it’s clear on its face. Here are some examples

(a) Are estimation and forecasting the same thing? As far as software estimation is concerned, yes they are. (Just do a Google or Bing search of “definition of forecast”.) Estimation, forecasting, prediction–it’s all the same basic activity, as far as software estimation is concerned.

(b) Is showing someone several pictures of kitchen remodels that have been completed for $30,000 and implying that the next kitchen remodel can be completed for $30,000 estimation? Yes, it is. That’s an implementation of a technique called Reference Class Forecasting.

(c) Is doing a few iterations, calculating team velocity, and then using that empirical velocity data to project a completion date count as estimation? Yes it does. Not only is it estimation, it is a really effective form of estimation. I’ve heard people argue that because velocity is empirically based, it isn’t estimation. That argument is incorrect and shows a lack of basic understanding of the nature of estimation.

(d) Is estimation time consuming and a waste of time? One of the most common symptoms of lack of estimation skill is spending too much time on the wrong activities. This work is often well-intentioned, but it’s common to see well-intentioned people doing more work than they need to get worse answers than they could be getting.

4. Being able to estimate effectively is a skill that any true software professional needs to develop, even if they don’t need it on every project. 

“Estimating is problematic, therefore software professionals should not develop estimation skill” – this is a common line of reasoning in #NoEstimates. Unless a person wants to argue that the need for estimation is rare, this argument is not supported by the rest of #NoEstimate’s premises.

If I agreed, for sake of argument, that 50% of the projects don’t need to be estimated, the other 50% of the projects would still benefit from the estimators having good estimation skills. If you’re a true software professional, you should develop estimation skill so that you can estimate competently on the 50% of projects that do require estimation.

In practice, I think the number of projects that need estimates is much higher than 50%.

5. Estimates serve numerous legitimate, important business purposes.

Estimates are used by businesses in numerous ways, including:

  • Allocating budgets to projects (i.e., estimating the effort and budget of each project)
  • Making cost/benefit decisions at the project/product level, which is based on cost (software estimate) and benefit (defined feature set)
  • Deciding which projects get funded and which do not, which is often based on cost/benefit
  • Deciding which projects get funded this year vs. next year, which is often based on estimates of which projects will finish this year
  • Deciding which projects will be funded from CapEx budget and which will be funded from OpEx budget, which is based on estimates of total project effort, i.e., budget
  • Allocating staff to specific projects, i.e., estimates of how many total staff will be needed on each project
  • Allocating staff within a project to different component teams or feature teams, which is based on estimates of scope of each component or feature area
  • Allocating staff to non-project work streams (e.g., budget for a product support group, which is based on estimates for the amount of support work needed)
  • Making commitments to internal business partners (based on projects’ estimated availability dates)
  • Making commitments to the marketplace (based on estimated release dates)
  • Forecasting financials (based on when software capabilities will be completed and revenue or savings can be booked against them)
  • Tracking project progress (comparing actual progress to planned (estimated) progress)
  • Planning when staff will be available to start the next project (by estimating when staff will finish working on the current project)
  • Prioritizing specific features on a cost/benefit basis (where cost is an estimate of development effort)

These are just a subset of the many legitimate reasons that businesses request estimates from their software teams. I would be very interested to hear how #NoEstimates advocates suggest that a business would operate if you remove the ability to use estimates for each of these purposes.

The #NoEstimates response to these business needs is typically of the form, “Estimates are inaccurate and therefore not useful for these purposes” rather than, “The business doesn’t need estimates for these purposes.”

That argument really just says that businesses are currently operating on much worse quality information than they should be, and probably making poorer decisions as a result, because the software staff are not providing very good estimates. If software staff provided more accurate estimates, the business would make better decisions in each of these areas, which would make the business stronger.

This all supports my point that improved estimation skill should be part of the definition of being a true software professional.

6. Part of being an effective estimator is understanding that different estimation techniques should be used for different kinds of estimates. 

One thread that runs throughout the #NoEstimates discussions is lack of clarity about whether we’re estimating before the project starts, very early in the project, or after the project is underway. The conversation is also unclear about whether the estimates are project-level estimates, task-level estimates, sprint-level estimates, or some combination. Some of the comments imply ineffective attempts to combine kinds of estimates—the most common confusion I’ve read is trying to use task-level estimates to estimate a whole project, which is another example of lack of software estimation skill.

Effective estimation requires that the right kind of technique be applied to each different kind of estimate. Learning when to use each technique, as well as learning each technique, requires some professional skills development.

7. Estimation and planning are not the same thing, and you can estimate things that you can’t plan. 

Many of the examples given in support of #NoEstimates are actually indictments of overly detailed waterfall planning, not estimation. The simple way to understand the distinction is to remember that planning is about “how” and estimation is about “how much.”

Can I “estimate” a chess game, if by “estimate” I mean how each piece will move throughout the game? No, because that isn’t estimation; it’s planning; it’s “how.”

Can I estimate a chess game in the sense of “how much”? Sure. I can collect historical data on the length of chess games and know both the average length and the variation around that average and predict the length of a game.

More to the point, estimating software projects is not analogous to estimating one chess game. It’s analogous to estimating a series of chess games. People who are not skilled in estimation often assume it’s more difficult to estimate a series of games than to estimate an individual game, but estimating the series is actually easier. Indeed, the more chess games in the set, the more accurately we can estimate the set, once you understand the math involved.

8. You can estimate what you don’t know, up to a point. 

In addition to estimating “how much,” you can also estimate “how uncertain.” In the #NoEstimates discussions, people throw out lots of examples along the lines of, “My project was doing unprecedented work in Area X, and therefore it was impossible to estimate the whole project.” That isn’t really true. What you would end up with in cases like that is high variability in your estimate for Area X, and a common estimation mistake would be letting X’s uncertainty apply to the whole project rather than constraining it’s uncertainty just to Area X.

Most projects contain a mix of precedented and unprecedented work, or certain and uncertain work. Decomposing the work, estimating uncertainty in different areas, and building up an overall estimate from that is one way of dealing with uncertainty in estimates.

9. Both estimation and control are needed to achieve predictability. 

Much of the writing on Agile development emphasizes project control over project estimation. I actually agree that project control is more powerful than project estimation, however, effective estimation usually plays an essential role in achieving effective control.

To put this in Agile Manifesto-like terms:

We have come to value project control over project estimation,
as a means of achieving predictability

As in the Agile Manifesto, we value both terms, which means we still value the term on the right.

#NoEstimates seems to pay lip service to both terms, but the emphasis from the hashtag onward is really about discarding the term on the right. This is a case where I believe the right answer is both/and, not either/or.

10. People use the word “estimate” sloppily. 

No doubt. Lack of understanding of estimation is not limited to people tweeting about #NoEstimates. Business partners often use the word “estimate” to refer to what would more properly be called a “planning target” or “commitment.” Further, one common mistake software professionals make is trying to create estimates when the business is really asking for a commitment, or asking for a plan to meet a target, but using the word “estimate” to ask for that.

We have worked with many companies to achieve organizational clarity about estimates, targets, and commitments. Clarifying these terms makes a huge difference in the dynamics around creating, presenting, and using software estimates effectively.

11. Good project-level estimation depends on good requirements, and average requirements skills are about as bad as average estimation skills. 

A common refrain in Agile development is “It’s impossible to get good requirements,” and that statement has has never been true. I agree that it’s impossible to get perfect requirements, but that isn’t the same thing as getting good requirements. I would agree that “It is impossible to get good requirements if you don’t have very good requirement skills,” and in my experience that is a common case.  I would also agree that “Projects usually don’t have very good requirements,” as an empirical observation—but not as a normative statement that we should accept as inevitable.

Like estimation skill, requirements skill is something that any true software professional should develop, and the state of the art in requirements at this time is far too advanced for even really smart people to invent everything they need to know on their own. Like estimation skill, a person is not going to learn adequate requirements skills by reading blog entries or watching short YouTube videos. Acquiring skill in requirements requires focused, book-length self-study or explicit training or both.

Why would we care about getting good requirements if we’re Agile? Isn’t trying to get good requirements just waterfall? The answer is both yes and no. You can’t achieve good predictability of the combination of cost, schedule, and functionality if you don’t have a good definition of functionality. If your business truly doesn’t care about predictability (and some truly don’t), then letting your requirements emerge over the course of the project can be a good fit for business needs. But if your business does care about predictability, you should develop the skill to get good requirements, and then you should actually do the work to get them. You can still do the rest of the project using by-the-book Scrum, and then you’ll get the benefits of both good requirements and Scrum.

12. The typical estimation context involves moderate volatility and a moderate levels of unknowns

Ron Jeffries writes, “It is conventional to behave as if all decent projects have mostly known requirements, low volatility, understood technology, …, and are therefore capable of being more or less readily estimated by following your favorite book.”

I don’t know who said that, but it wasn’t me, and I agree with Ron that that statement doesn’t describe most of the projects that I have seen.

I think it would be more true to say, “The typical software project has requirements that are knowable in principle, but that are mostly unknown in practice due to insufficient requirements skills; low volatility in most areas with high volatility in selected areas; and technology that tends to be either mostly leading edge or mostly mature; …; and are therefore amenable to having both effective requirements work and effective estimation work performed on those projects, given sufficient training in both skill sets.”

In other words, software projects are challenging, and they’re even more challenging if you don’t have the skills needed to work on them. If you have developed the right skills, the projects will still be challenging, but you’ll be able to overcome most of the challenges or all of them.

Of course there is a small percentage of projects that do have truly unknowable requirements and across-the-board volatility. I consider those to be corner cases. It’s good to explore corner cases, but also good not to lose sight of which cases are most common.

13. Responding to change over following a plan does not imply not having a plan. 

It’s amazing that in 2015 we’re still debating this point. Many of the #NoEstimates comments literally emphasize not having a plan, i.e., treating 100% of the project as emergent. They advocate a process—typically Scrum—but no plan beyond instantiating Scrum.

According to the Agile Manifesto, while agile is supposed to value responding to change, it also is supposed to value following a plan. Doing no planning at all is not only inconsistent with the Agile Manifesto, it also wastes some of Scrum’s capabilities. One of the amazingly powerful aspects of Scrum is that it gives you the ability to respond to change; and that doesn’t imply that you need to avoid committing to plans in the first place.

My company and I have seen Agile adoptions shut down in some companies because an Agile team is unwilling to commit to requirements up front or refuses to estimate up front. As a strategy, that’s just dumb. If you fight your business up front about providing estimates, even if you win the argument that day, you will still get knocked down a peg in the business’s eyes.

Instead, use your velocity to estimate how much work you can do over the course of a project, and commit to a product backlog based on your demonstrated capacity for work. Your business will like that. Then, later, when your business changes its mind—which it probably will—you’ll be able to respond to change. Your business will like that even more. Wouldn’t you rather look good twice than look bad once?

14. Scrum provides better support for estimation than waterfall ever did, and there does not have to be a trade off between agility and predictability. 

Some of the #NoEstimates discussion seems to interpret challenges to #NoEstimates as challenges to the entire ecosystem of Agile practices, especially Scrum. Many of the comments imply that predictability comes at the expense of agility. The examples cited to support that are mostly examples of unskilled misapplications of estimation practices, so I see them as additional examples of people not understanding estimation very well.

The idea that we have to trade off agility to achieve predictability is a false trade off. In particular, if no one had ever uttered the word “agile,” I would still want to use Scrum because of its support for estimation and predictability.

The combination of story pointing, product backlog, velocity calculation, short iterations, just-in-time sprint planning, and timely retrospectives after each sprint creates a nearly perfect context for effective estimation. Scrum provides better support for estimation than waterfall ever did.

If a company truly is operating in a high uncertainty environment, Scrum can be an effective approach. In the more typical case in which a company is operating in a moderate uncertainty environment, Scrum is well-equipped to deal with the moderate level of uncertainty and provide high predictability (e.g., estimation) at the same time.

15. There are contexts where estimates provide little value. 

I don’t estimate how long it will take me to eat dinner, because I know I’m going to eat dinner regardless of what the estimate says. If I have a defect that keeps taking down my production system, the business doesn’t need an estimate for that because the issue needs to get fixed whether it takes an hour, a day, or a week.

The most common context I see where estimates are not done on an ongoing basis and truly provide little business value is online contexts, especially mobile, where the cycle times are measured in days or shorter, the business context is highly volatile, and the mission truly is, “Always do the next most useful thing with the resources available.”

In both these examples, however, there is a point on the scale at which estimates become valuable. If the work on the production system stretches into weeks or months, the business is going to want and need an estimate. As the mobile app matures from one person working for a few days to a team of people working for a few weeks, with more customers depending on specific functionality, the business is going to want more estimates. Enjoy the #NoEstimates context while it lasts; don’t assume that it will last forever.

16. This is not religion. We need to get more technical and economic about software discussions. 

I’ve seen #NoEstimates advocates treat these questions of requirements volatility, estimation effectiveness, and supposed tradeoffs between agility and predictability as value-laden moral discussions in which their experience with usually-bad requirements and usually-bad estimates calls for an iterative approach like pure Scrum, rather than a front-loaded approach like Scrum with a pre-populated product backlog. In these discussions, “Waterfall” is used as an invective, where the tone of the argument is often more moral than economic. That religion isn’t unique to Agile advocates, and I’ve seen just as much religion on the non-Agile sides of various discussions. I’ve appreciated my most recent discussion with Ron Jeffries because he hasn’t done that. It would be better for the industry at large if people could stay more technical and economic more often.

For my part, software is not religion, and the ratio of work done up front on a software project is not a moral issue. If we assume professional-level skills in agile practices, requirements, and estimation, the decision about how much work to do up front should be an economic decision based on cost of change and value of predictability. If the environment is volatile enough, then it’s a bad economic decision to do lots of up front requirements work just to have a high percentage of requirements spoil before they can be implemented. If there’s little or no business value created by predictability, that also suggests that emphasizing up front estimation work would be a bad economic decision.

On the other hand, if the business does value predictability, then how we support that predictability should also be an economic decision. If we do a lot of the requirements work up front, and some requirements spoil, but most do not, and that supports improved predictability, and the business derives value from that, that would be a good economic choice.

The economics of these decisions are affected by the skills of the people involved. If my team is great at Scrum but poor at estimation and requirements, the economics of up front vs. emergent will tilt one way. If my team is great at estimation and requirements but poor at Scrum, the economic might tilt the other way.

Of course, skill sets are not divinely dictated or cast in stone; they can be improved through focused self-study and training. So we can treat the question of whether we should invest in developing additional skills as an economic issue too.

What is the cost of training staff to reach proficiency in estimation and requirements? Does the cost of achieving proficiency exceed the likely benefits that would derive from proficiency? That goes back to the question of how much the business values predictability. If the business truly places no value on predictability, there’s won’t be any ROI from training staff in practices that support predictability. But I do not see that as the typical case.

My company and I can train software professionals to become proficient in both requirements and estimation in about a week. In my experience most businesses place enough value on predictability that investing a week to make that option available provides a good ROI to the business. Note: this is about making the option available, not necessarily exercising the option on every project.

My company and I can also train software professionals to become proficient in a full complement of Scrum and other Agile technical practices in about a week. That produces a good ROI too. In any given case, I would recommend both sets of training. If I had to recommend only one or the other, sometimes I would recommend starting with the Agile practices. But I wouldn’t recommend stopping with them.

Skills development in practices that support predictability vs. practices that support agility is not an either/or decision. A truly agile business would be able to be flexible when needed, or predictable when needed. A true software professional will be most effective when skilled in both skill sets.

17. Agility plus predictability is better than agility alone. 

If you think your business values agility only, ask your business what it values. Businesses vary, and you might work in a business that truly does value agility over predictability or that values agility exclusively.

In some cases, businesses will value predictability over agility. Odds are that your business actually values both agility and predictability. The point is, ask the business, don’t just assume it’s one or the other.

I think it’s self-evident that a business that has both agility and predictability will outperform a business that has agility only. We need to get past the either/or thinking that limits us to one set of skills or the other and embrace both/and thinking that leads us to develop the full set of skills needed to become true software professionals.