"What is probabilistic forecasting (and why is it better)?" Find out in this episode of Appfire Presents: The Best Project Portfolio Management Show by Appfire. Julia Wester talks with Appfire's own Kerry O'Shea Gorgone about probabilistic forecasting - what it is, why it's better for project outcomes, and how it works.
About the guest
Julia Wester is a co-founder of 55 Degrees, an outcome-focused product company. Julia's the product owner for their flagship product, ActionableAgile.
About the show
The BEST Project Portfolio Management Show by Appfire covers everything you ever wanted to know about PPM by talking with project management experts who’ve seen it all. And every episode is 10 minutes or less, so you can get back to changing the world, one project at a time.
For your convenience, here is the transcript of this episode:
What is probabilistic forecasting (and why is it better)?
Kerry: Today we’re going to talk about probabilistic forecasting, what it is, why it’s better for project outcomes, and how it works. Julia Wester is going to join us. She is co-founder of 55 Degrees, an outcome focused product company, and she is the product owner for the company’s flagship product Actionable Agile. Stick around because you’re not going to want to miss the next 10 minutes.
Julia, thanks for joining us. Probabilistic forecasting, how is it different from what we normally do when we try to estimate how long a project is going to take, or how much effort, or when we might be finished?
Julia: It has a sort of scary sounding name, but probabilistic just means that there are multiple possible outcomes. So, what’s the probability that you’ll have a specific outcome or not?
Probabilistic forecasting is the way that you want to forecast when there is more than one possible outcome. In other words, we can’t be certain of exactly what’s going to happen. Let’s face it, that’s all the time for us. We very rarely know exactly what’s going to happen down to the letter.
When I’m giving someone a probabilistic forecast and teaching people what they are, there’s two pieces. Once you know these two pieces, you can start recognizing them out in the wild.
Every probabilistic forecast has a probability – that’s in the name, so it’s a given – the odds, the chances that something that will happen. Like there’s 100% chance of rain or a 55% chance of rain.
Then there’s a range. There’s a 50% chance that we can finish an item in 10 days or less. So, I have the 85% or 50%, or whatever our percent needs to be, that’s one way of showing risk. Then the range of possible outcomes, 10 days or less. I’m not saying things will take exactly 10 days. I’m saying it will take anything up to 10 days.
Another way is something like 75% chance we’ll finish 15 or more items in a sprint. If you hear those two things together, you’re hearing a probabilistic forecast.
Kerry: Okay. It’s like we have Atlassian Summit coming up, what’s the chances I’m going to have this booth ready? Like that?
Julia: Yes. You can think about it like that. Generally, we’re looking at forecasting how long it takes to finish given work items. You could say, “Given our past data on prepping for summits or things like that, what’s the likelihood that we’ll get all of our work done by the drop-dead date for that?” You could totally use the tool.
Kerry: What do you need? You were just touching on a couple of data points it sounds like it would be useful for me to have to know if I should bother showing up for Summit or not.
Julia: All the data that you need is you just need to have completed some work. You don’t need to have super fancy curated data. You need to have started and finished work items. Literally, that means for each item you need a start date and a finish date. That’s all you’re going to need for probabilistic forecasting.
You need data that is generated, and I mean finished items. Here’s a tip. We’re going to use our past data to help us forecast what will happen in the future. Getting to the why is it better, because that’s so fast. I don’t have to spend hours, days, and weeks estimating.
In order to have that actually work for us, we need to have data that has been generated, work that has been finished under roughly similar conditions. Meaning the situation I had when I finished that work I’m using needs to be roughly similar to the conditions I’ll have for the time I’m trying to forecast.
Kerry: So, if you lost half your team or something in the Great Resignation, you would need to factor that in.
Julia: You would. You could still use this kind of forecast with a grain of salt. You might also want to try another method of forecast and see what’s different. But all that while you’re generating new data.
One of the things that people ask me that goes under “what do I need,” is how much data do I need to do that. Do I need years of data? People are really surprised how quickly you find the boundaries of what is likely to happen. You can get a good forecast with just 10 finished items.
But… There’s a but. But ideally that set of items that you have represents the mix of work that you’re trying to forecast. You can’t be frontloading it with all of one type and then be trying to forecast lots of different types of work. The path needs to look a little bit like what your future is going to look like.
Kerry: Compare apples to apples.
Julia: Right. Going back to the conditions. The team conditions shouldn’t have had any drastic changes. We’re not trying to be super specific, like we had one more day off this time than we had in the past. This is at a rough scale we want to be approximately right instead of exactly wrong. Precision isn’t necessary.
Kerry: But if it’s December when you’re doing a project, that’s different than if you’re doing it in May, because of all the time off people take and that kind of thing. So, you need that data at hand.
Julia: Yes. There are some times where it might feel a little bit difficult to find exactly the right data. In that case, you can look back and see is there another time period which was similar or is that too far back to really be representative anymore. Again, in that case when you can’t be sure that you have a generally representative set of data, you can use this as one indicator. You can also use a more expensive way of estimating in that case, just to see if it comes up with a different outcome.
If you’re considering changing the way you forecast and moving to probabilistic forecasting, you either need to have one of two things. You need to either have the same level of accuracy with a lot less effort or cost, and/or you need a better level of accuracy. If you land on either of those two fronts, why not do that as well? In fact, we suggest when you’re getting started, continue with your current way you estimate and forecast and then try this alongside and see for yourself what works better for you, what’s less costly for you.
You’ll find along the way that there are things that you might be able to do about the way that you work to make your system a bit more stable and to make your data better at giving you forecasts. In this way, starting to do probabilistic forecasting actually drives you to do better work and better processes.
Kerry: What are some things you can do then to make this forecasting easier and more accurate?
Julia: This goes into more predictable data is better for forecasting, so what we’re really asking is how can we be more predictable. The very best thing that you can do, if you’re only going to do one thing, pay attention to how old the work that’s in progress is. Your work in progress, the stuff that’s already started. Don’t worry about the stuff yet that you haven’t started.
Try to minimize the age, the time it takes for you to finish something. Because we don’t know how old our work is, it’s hard to see that no matter what you’re using. Once you start realizing it, that starts being an input into your decision making. As we do things, like limit the amount of work we have in progress and work on our dependencies and do things that make age less, then we have better data coming out of our system, we’re delivering things more regularly, they’re taking less time. Those are the conditions that make for really good probabilistic forecasting. It’s like you want to be good at this because then you’ve done a lot of other things right.
Kerry: Do I need to have an app to do probabilistic forecasting, some software or something, or can I do it by hand?
Julia: You can absolutely do all of this without buying anything. In fact, a lot of it you can do with just a pencil and paper. There are two types of forecasts. Forecasting single items, and all that requires is knowing how long it took you to finish single items in the past. You can plot that all down. You can reach out to me at Julia@55degrees.se and I can tell you how to do that by hand.
Multiple item forecasting, looking at rates at which we finish work, requires more of a simulation. There are Excel spreadsheets floating around out there as well, but there are apps, like our app Actionable Agile for Jira, on the marketplace that we think are integrated into your work systems already that can just make your work life a lot easier without having to do all of that extra work. You can use these tools – we’re not the only ones, but we think we’re the best…
Kerry: Like this is The Best Project Portfolio Management Show by Appfire.
Julia: And if we’re not yet, we’re trying to be. They literally let you get risk aware forecasts in minutes. Try us out. You can find us on the Atlassian Marketplace and at 55degrees.se.
Kerry: And you can find The Best Project Portfolio Management Show at Appfire.com, so be sure to come back for more 10-minute episodes where we give you everything you need to be awesome at project management.
Julia, thanks for joining us.
Julia: Thank you.
Kerry: Thanks to everyone who watched and listened.