005: Behind The Predictive Analytics Curtain w/ Gauthier Vasseur

Ever since, I’ve always looked for data in my jobs, whether it was finance, treasury or audit. I rapidly found out that when you master data, when you know how to process it properly and efficiently, you gain a lot of time. You actually get on top of things faster. You get more comfortable, which gives you the ability to do more and progress in your life and in your career.

The last thing I wanted is to see them wasting their timing on useless tasks, manual tasks. I really thought that this has to be changed. I started to teach, I started to write articles, I started to share about all the things I had learned on data to make sure my employees, my interns, my students would be able to take it from there and then spread the gospel in turn.

All these disciplines are journey, you never see the end of them, there is always something you can learn, you can do better. There is always a new technique that you are looking forward to mastering. I think that’s all the people I’ve seen around me being successful at data or astro-photography, all know they are on the journey and they never stopped learning.

That’s why I was very happy to participate to this webcast because I think it’s part of this continuous learning we all have to do on this stuff to doing more or getting better with data.

When you talk about learning. There is a new, I would say, how do I put it? There is a new interest in predicting future actions based on past events. That’s where you talk a lot about predictive analytics.

For the folks who’ve been in this space similar to me fifteen years or more, where you are still used to the old ways of doing things, the old way of packaging data. How does predictive analytics fit into this journey of moving forward to the future with new tools and memory computing, how do we adapt to this new way of consuming and packaging data?

Finally, we can achieve what science used to do for research. We can achieve that for domains such as accounts receivable, for domains such as stock analysis. If we go back a little bit, predictive analytics always sounds like a big buzz word. Overall it doesn’t have to be that complex.

The thing is, these predictions, the way they bring value is, as you consistently create this prediction, you’ll find out that they will beat your hunches. They will be better than the crystal ball. You’ll find out that even if they don’t have to be perfectly right, they will always be slightly better than anything you have.

Predictive analytics for me is both math and statistics and it’s also an approach to solving problems that if applied consistently will end up being positive overall.

At the same time they are not magical, it’s not looking in the future, it’s taking the best we have, and the best math and statistics to predict something.

The example that Nick Silver brings up is predicting where a hurricane is going to hit. We know there are a lot of hurricanes in the Mexican gulf hitting Texas, Florida. Twenty years ago, the prediction level was pretty much a 350 miles radius, which means, whenever a hurricane was coming up the coast, the local authorities had a 350, miles radius margin of area to evacuate.

Which means any time a hurricane would hit, you’d say, “Evacuate all these coastlines,” which are millions and millions of people. With a better predictive tool you can say, you can predict accuracy at a hundred mile radius, which still sucks a little a hundred miles, come on.

Guess what? A hundred miles radius narrows down the number of people, the zone to evacuate to a big metropolitan area. We are not talking about entire state coastlines. We are talking about large cities. That completely changes the game in how to evacuate people and work two focused resources in order to prevent big disasters.

The way you described it as math and statistics is less scary, is more concrete, is more grounded in science. The takeaway that I’ve had from you is, even though it’s not a perfect guess or estimation of the future, you are really connecting the dots to making more accurate, what’s a better word for prediction?

How does a predictive tool help them to see one month, three months into the future for mundane tasks such as demand planning or forecasting?

To build these predictive models, companies are going to need tremendous amounts of data and not the aggregating kind. They need it at the lowest granular level, with every single little detail. That’s what statistical model we’ll use to derive this prediction.

You mentioned earlier the notion of black box. I think for all of us, the different steps are quite simple but there is indeed within this other creation of this prediction. There is a bit of heavy duty science that is brought in. That’s why we all dream of doing predictive analytics.

Unless we are PhD mathematicians, we might find ourselves a bit short. That’s why pieces of software, that’s why software is there is willing to help us. That’s why applications are designed.

The first one is gather the data. You need to secure that data because you need a sample to generate your predictions from.

Once you have your data you need to be able to articulate this information in meaningful business format because stat is stat but data is data. If you don’t know what the data means, statistics won’t know much more than you do.

You need to get the business being the business orientation to that data and then apply the statistic correctly on the data.

We talked about the data but business data has some very strange specificities and you need to use robust statistics and this is a great pickup line if you want to shine at business dinners. The concept of maximum entropy. These are concepts you need to apply to your predictive models so that outliers in your business data don’t influence too much so that the results of the prediction can really be reliable.

That is the layer of black-boxing you absolutely need to have in statistics. At the end of the day, we all agree, this is terrific to predict demand, to predict machine breaks down, inventory level, but you still need the people to be able to understand that data and interact with it. There is a need for strong UI, strong graphical capabilities to make all that work.

Does a predictive tool like Trufa bridge that gap for business users who don’t have a PhD in statistics to at least apply some rudimentary models against that data?

Once you have that, then we can go back to simple concepts. All it takes is pick that data up, bring it to a place where it’s going to be organized, where every data is going to be interpreted the way it should be. Every piece of information is going to be put part of a process of the transactional process.

Then you let the calculation power, the math power, the statistical power get started and you display all this in a UI, a user experience that guides any one of us, business people through different steps where you are pointed to opportunities, you are guided through simulation and then you can really predict what your actions will create in terms of economic impact.

If you recall decades ago, the concept of the efficient frontier was created in the world of stock management, investment management, where for a given risk you would assess how much yield you could have. The frontier would be actually the perfect curve where you have the maximum yield for the minimum risk.

What it shows you is that you can strike balance between all the drivers that define performance. To give you an example, a very classic one. Actually I’ve run into that very case when I was in treasury and I could never solve it.

You want to reduce working capital. You are going to see what impacts working capital. You are going to look at rebates, you are going to look at yelling at your customers, yelling at your vendors, you might change your operational processes. You would see result in working capital.

Rebates have taken down revenue. Pushing customers have done great in quality of service and they are not renewing their commitments. Pressing vendors has degraded the quality of the pieces you are being served.

At the end of the day you push on one side of the scale and the other side went the other way. By identifying and measuring the impact of these drivers when they move, where there is some profitability or working capital. By doing it on both sides, you can define that efficient business frontier. I think that is the ultimate challenge in predictive analytics.

All the things that you pointed to, it sounds, like you said not a magic bullet but it does give a forward perspective on what you can do without spending too much time building and testing models to see if it works.

I think for those of us who now have a renewed interest in finding more about predictive analytics. Where can we find more information or innovations with that you are doing with predictive and where can we learn more about what the tool has to offer for business users who are just trying to maximize working capital, either in the treasury or at least in the space I’m interested in, logistics and supply chain?

The book I mentioned earlier The Signal and the Noise by Nick Silver is also a very easy read, perfect for the beach and gives you tons of great stories that will make you look good at dinners.

On a more serious standpoint, talking about what I do, definitely the trufa.net website, www.trufa.net website also contains many examples on how predictive analytics and the concept of efficient business frontier applies to organizations.

There are actually a couple of pages I love a lot. One is Guenther’s blog.

If you have a little bit more courage but still very easy read, two great papers from our chief scientist, Professor Andreas Mielke about maximum entropy and the concept of robust statistics.

Just the first page will really illuminate you on how simple the base concepts of predictive analytics can be. Then on the second page how heavy duty mathematics it’s all about.

That also tells you that, these things are great but let’s not forget, this is serious math. If it’s not as part of the integrated application, get ready for a lot of fun with some PhDs.