Microsoft Finance Suite’s AI/ML Journey: Part 2 – AN FP&A Podcast

On the latest FP&A Today podcast, sponsored by Datarails, the financial planning and analysis platform for Excel users, host Paul Barnhurst interviews Microsoft data scientist Daniel Sousa-Lenox.

In this, Part 2 of the FP&A Today podcast, Barnhurst and Sousa-Lenox examine why AI is not a “magic bullet” for finance, despite its success, along with two of the most powerful lessons in presenting AI to finance teams. And Apple Pod.

Check out Part 1 of our interview with Sousa-Lenox on our blog page.

For this and all of the FP&A podcasts, visit the website FPAToday podcast at Podcast Bean and at the Apple Podcast website.

Paul Barnhurst: I love how you said machine learning is not the magic bullet, but for certain use cases, in many situations, it can be highly accurate and it saves a ton of time for finance professionals. There’s an organization called FP&A Trends and we had the lead analyst on the podcast recently.

The company has an annual survey that tracks how well finance pros thought of their budget and forecasting systems. 39% said their budget is good to great while 16% said their budget process was basically useless across all boards. As you went up and looked at those who were using AI and machine learning, the survey found the first number went from 39 to 63% in calling their forecast “good or great”. That shows using technologies with the human touch can vastly improve the budget forecasting process.

Daniel Sousa-Lennox: Yes, yes, definitely.

Paul Barnhurst: It sounds like that’s what you’re seeing at Microsoft.

Daniel Sousa-Lennox: Yes, for sure.

Paul Barnhurst: I saw your boss last year at the AFP who gave a fascinating presentation about Microsoft and how they were using machine learning. When you talked about the algorithm you guys were using called FINN, and how they had made the technology open source. Can you tell our audience a little bit about FINN, what it is, and how people can use it?

Daniel Sousa-Lennox: At Microsoft, we have a few teams that have completely adopted the use of machine learning specifically with our tool called the Microsoft Finance Time Series Forecasting Framework (FINN) an automated forecasting framework for producing financial forecasts.

Basically, FINN is a standardized modeling framework we built for our team that looks to allow anyone in finance or actually anyone beyond finance to use the technology. That said, we’re targeting specifically our internal finance audience. FINN allows any finance team member to produce a machine learning forecast with no need to code anything or to learn what an arena model is and how to code one in Python. There’s just none of that.

Internally at Microsoft, we built a UI, similar to a web UI, where the user can interact with them through dropdown menus. So it’s absolutely a “no code” tool internally.

Externally, as you mentioned, we’ve made it an open-source package that was built in R, which is one of the two most popular languages for ML, probably after Python. So we built it in R and we made it open source, which means anyone can go into GitHub into the Microsoft system and look for FINN.

Paul Barnhurst: What does FINN do best?

Daniel Sousa-Lennox: What FINN does is automate all the tedious processes of doing machine learning, which means doing the data preparation first and making sure it’s how it needs to be for the models to ingest the data. That includes cleaning the data, analyzing if you have outliers, and taking care of them.

You can also use the technology for feature engineering, which is basically having a set amount of columns that you have in your dataset and adding more information to it by decomposing such columns, like a calendar date for example.

You just separate the day from the month which gives valuable information to models. FINN automates that part as well. Finance professionals can also conduct backtesting to see if the model is accurate.

Lastly, FINN creates your future forecast, along with how far into the future you want your forecast to be. We can do daily all the way to yearly and everything in between. All you need to do as a user is bring historical data, with as high quality as possible and as much historical data as possible, but that’s really the only requirement and FINN will do the rest. All in about 90 minutes instead of days or even weeks.

Paul Barnhurst: It sounds like FINN helps you set up and clean up the data in the sense of getting it in the right format for running all the algorithms and running the different methods. Then it compares each of those based on criteria, whether it’s correlation, Z-score, or whatever, and lets you know the best way to use the data. But does it provide you with the top recommendation, and then the second and third recommendation, or does it just give you one recommendation?

Daniel Sousa-Lennox: Yes, the output gives you the forecast for every model that FINN has in it, but it checks with a checkbox on which forecast was the best-performing one.

That way, you get access to all of the forecasts in case you do want to go deeper and understand what one model does, and maybe for some reason you do want to go a different route, you can do that.

You get access to all those results. Yet FINN clearly states this model or this combination of models were the ones that led to the highest accuracy in your forecast and therefore these are the future values that we suggest you use.

Paul Barnhurst: Obviously you work closely with FP&A to improve the forecasting models.

So talk to us a little bit about how that process typically goes. You mentioned earlier black box, as sometimes not being able to trust. How can you work with them and are there any challenges you find or how does that process go?

Daniel Sousa-Lennox: Yes, I think the first thing, when the team is ready to try machine learning, the first thing is creating that clarity that, as we said previously, emphasizes the technology is not “magic”.

Since ML is not a crystal ball that looks into the future, the first objective as a starting point is to get something decent at first. That’s one.

The second one is also creating clarity. That means having an iterative process that takes a lot of stages because, at first, you get a result and it’s not as accurate as you want. You need to go back. You need to think of what impacts my business that I could add to the data. The other thing that happens a lot is people usually just bring in the historical data of whatever metrics you want to predict, and that’s it.

So if you want to predict revenue for your company three months into the future, all you need to do is bring your revenue for the past three years. That could be useful if you have a business that has a high correlation or auto correlation, meaning what happened a year ago significantly impacts what’s going to happen the next year. That usually happens with a subscription business, but in general, you need to know what drives your business, what metrics of the economy, or what company metrics or industry metrics impact your products.

So that’s one thing. Getting the teams to really think through those issues is very important.

Additionally, even as you could go as high up or as granular as you want, there’s possibly a level in between that’s going to hit that sweet spot of accuracy. Only you as the finance person or the product owner or the service are the ones best suited to help translate that knowledge into data. You can simply toggle this input and maybe change this when you’re setting up FINN, but 80% of it or more is going to come from what expertise you have and your ability to translate that into data to be ingested by the models.

Paul Barnhurst: If you could offer advice to someone starting a career in FP&A today, what advice would you give them?

Daniel Sousa-Lennox: The best advice would be to fight that urgency mode that leads you to just going through the motions and updating spreadsheets and just doing the task at hand without going through that analysis portion. I know it’s going to be extremely hard because things are due now or even yesterday in this day and age. That said, emphasizing analysis is how you’re going to actually grow within your organization in an FP&A role.

Because getting those numbers crunched faster, yes, that’s noticeable and that’s impactful and that’s helpful, but in the end, that’s not going to leave the same mark that it would if you have those insights. If you can bring the numbers crunch to your leadership team, but alongside those, you already have insights, you already have ideas, you already have recommendations, that’s how you’re going to stand out, more so than if you just bring really good, nice quick numbers.

If you can anticipate those requests get those insights and bring those up to your team and your leadership team, that’s going to make a significant difference both in your career and how much you get out of it.



Recent Posts

Leave a Reply

Your email address will not be published. Required fields are marked *