How Can Artificial Intelligence Fuel Your Cash Forecasting Accuracy: Case Study On distribution now

[0:01] Host Speaker:

Hello, everyone. Thank you for joining today’s session on how AI can help the oil and gas industries to forecast their cash better. We have with us Narender Nimmagadda and Amber Thompson. Narender is the Director of Product Management Treasury at Highradius and Amber is a functional consultant Treasury at Highradius. To make the sessions more interactive raise your hand if you have a question. And a mic will be brought to you so everyone is able to hear your question. Please state your name and organization first and then ask your question. We’re also reserving the final 10 minutes of the session for open q&a and discussion. Post-session materials will be available in the hover app. At the end of this session, we ask that you quickly use the app to rate the presentation. This helps assure we are continuously bringing you the best content that is relevant to you. Also, I’ll be circulating this sheet among you people so that you can enter your name and email address so that we can give you the presentation slides and you will be able to gain one credit for CTP recertification. So please fill that up. Thank you. So now everything is sorted. Amber and Narender, the stage is all yours. Thank you. Round of applause everyone for them.

[1:14] Narender Nimmagadda:

Good morning guys. Hello, welcome to you all. Welcome to the radiance 2020 treasury’s session on how artificial intelligence can fuel your forecasting process. Okay, so I’m Narender from product management, working for Highradius into the treasury area. And today’s session, I’m presenting on behalf of Sansa, the VP of corporate treasury and development in distribution now. Today’s session will be a case study into what are the challenges for distribution now and how we have tried to solve their forecasting process. So the agenda that we have today is we’ll first have a quick glance into the business of distribution. Now, their current forecasting process, their challenges and the limitations in their forecast process and followed by a presentation on how the solution approach that we have taken to solve their forecasting process will be presented by my colleague. So first of all, about distribution now. So distribution now is a market leader in supplying to the energy markets across the world. They have around 250 locations spread across 20 countries. They’re a publicly listed company with around 3.1 billion. Their current forecast process is again typically a manual process. Today we are focusing on the A/R so in the AR they have three business divisions where the analysts collect data from these 3 business divisions, put them in excel sheets, aggregate them for some analysis, and process it and send it to the executive team. So focusing a little more on the air process. So what they do today in the current process is they take their current, open invoices and the collections data and on top of that, the process using the logic of average days to pay. Any delays or something like that, and the promises free information, and they use that information to process open invoices, data and all of these they’re collecting in the excel invoices. And then they aggregate this within the excel invoices, send it to the treasurer who validates them and then send it to the executive team. So there’s a typical process that they are following which is just very manual oriented and you know Excel oriented which we in our terms call it as Excel hell.

[4:08] Narender Nimmagadda:

So, their uncertainty in the area of financial AR forecasting as follows the oil prices which kind of affect their down the line revenues, the collections and everything. And apart from that, in some cases, the end customer behavior, disputes, rebates, and discounts, etc. So, these are the cases, which kind of cause the uncertainties in analyzing and estimating their forecasts. Now, the challenges that they primarily face when they’re processing within the AR forecasting is one they have to collect this information from multiple teams– different business divisions, get them all sent to a centralized location in an Excel format, aggregate them usual use manual formulas, and then on top of it, use the human judgment to know the forecast and all. They currently do a three to four-week forecast and typically their forecast drops off. The accuracy drops off after two weeks or three weeks. For the first week or two, the focus is reasonably accurate, but after that, the forecast accuracy drops. The implications of all of this are again standard; you know problems with having to maintain a high low cash buffer. And most of the time they are doing this forecast analysis trying to use human judgment which is more error-prone and probably also having limited information etc. Another major challenge or limitation that they have is their focus is more short term forecast, meaning they just have three weeks to four weeks forecast. They do not have long term visibility like they cannot have a what’s going to be the focus at the end of this quarter like you have three months or six months forecast. They do not have long term visibility and all of that.

[6:15] Amber Thompson:

Yeah, so I am the consultant for the email and I will be walking through how we are using our AI-powered forecasting to turbocharge their cash forecasting. So as Naren walks you through, they have a very manual process right now when it comes to your consolidated disbursements and AR forecasts. We will be at the end state, fully automating across categories with Highradius. Right now we’re in phase one of the implementation. We started for about a month. right to the beginning of the year. And we are focusing on what Hassan confirmed was his most challenging cash flow category, which is AR. So we have already set up an out of the box model for him to forecast AR and are planning on tuning that and making that better as we continue through our implementation. The impact that we’re going to see when we are forecasting AR alone will be a higher efficiency for distribution now, as well as much higher confidence, especially when it comes to longer-term forecasting. And this will be due to eliminating Excel and that manual process, as well as having all the information in one place for their users. Now let’s get to the numbers. Right now again, one month into our project, we have been able to reduce variance by 26% from what they are currently forecasting today. We are still tuning or model planning on tuning and targeting and about 90% plus accuracy going forward in AR for the United States. This will provide distribution now with a standardized and more efficient forecasting process, as well as providing them with a faster and more informed decision with their cash processes.

[8:36] Amber Thompson:

Let me actually walk you through a demonstration of what we are providing to them today. Again, this will be just with AR and we have manipulated the data so that we’re not showing you the live data. You guys should be able to see this now. This is our forecast system for AR. You will be able to see forecasts for a period of up to six months. And you can choose to bucket this forecast Do you know Hassan is able to bucket this forecast on both, a weekly and a monthly basis? We also have the capability of showing the forecast versus the actual that you’re seeing on those buckets as well. So that you can see how confident you are going to be in the system. That would be on this card here and move over to the right. We have a variance that shows you the difference between that forecast and the actual as well as a variance in the percentage of the actual itself. Now the piece I was most excited to hear about when we were showing this to him for the first time was this forecast model. We have turned this item into sheets, and it’s basically an Excel on the web, where when you click that button, you can be brought over to this screen. This screen will show you our bucketed forecasts upcoming, and all done through AI. And then in any of these manual overrides, you can choose to add any information that we would not be able to get by historical data.

[10:17] Amber Thompson:

As an example, if you have an upcoming acquisition, this could be something that historically would not be able to forecast itself. Another example as a little bit of more information is that we can walk through is if you do have someone collecting on your customers, and one of them calls up and tells you that they’re going to make an unusual payment of $50 million per se. You can come in here and choose to type in that amount.

[11:01] Amber Thompson:

And it will take just a moment to adjust to that. And when you come back here, all you have to do is see your forecast update with that information. Click this include manual override, and you should be able to see that both of these inflows have changed. So, back to our PowerPoint here. And we can come back to where we’re currently at today.

[11:43] Audience Speaker:

The question was, I’m assuming it’s just a preference or a configuration. Configuration, but the buckets for cash inflows were weekly and monthly. That could be daily also, right? Just a setting.

[12:03] Narender Nimmagadda:

Yes. So this allows you to configure in any manner. For Ehsaan, it was that the current process was weekly. Right? So that’s the reason we had to configure their system at a weekly level. But the system allows, in fact, the system does the forecasting at a daily level. So it’s just a view that can be configured and it can be seen any which way you want. And the maximum focusing that we are doing now is for a period of six months into the future. And even the variance analysis that amber was showing was you can go back into the past six months and see the whole system has been performing. So as you keep forecasting, you pass the time, and then you want to understand how well the system has been doing. The variance analysis at the bottom allows you to go and see to give you a comparison of actual versus forecast data. They get down into the variance analysis at each individual cash flow category level as well. In the case of distribution, now, since we had only AR, you can be able to see the analysis of the variance at the A/R level. But once you have the entire system up and running, where you’re forecasting for different cash flow categories individually, you will be able to see the variance analysis right down to the cash flow category level, or even slice and dice at our individual company level in case you have multiple company codes or multiple different regions. And not just that, in the roadmap, we also plan to have this variance at each individual customer level. Like, typically, if you have a top set of customers, let’s say the top 10 customers and follow the rest. You want to understand, hey, where is the biggest variance happening, right? Is it in the AR or AP, in the corporate sales or retail sales or export and which customer? Right, you can go into all that level of detail as well.

[13:55] Audience Speaker:

Hi, Aziz. You mentioned that the AI is like an auto box ready to implement. Could you talk a little bit more about what really is the connection to? I think you mentioned a little bit. It is connected straight to like customer past customer transactions, or does it go back to when the sales are happening? How are you configuring it at the initial stage? And then you just talked about how you go back and you look at the variance, but who’s making the call? Is a company data analytics team? Or going back and saying, “Okay, well this customer is too unpredictable,” or is it like your team? I mean, just trying to get a good understanding.

[14:39] Narender Nimmagadda:

Okay. So, essentially, again, going back into the setup. So what we do at the beginning of the implementation phase is we try to take your historical data, like the last two years or three years, the more the better. Of your bank transactions, your invoices, both receivables and payables and all that. Typically the cash forecast that you want to forecast. We feed this data into the system on which our AI algorithm runs, and it tries to understand the behavior patterns. It tries to understand the seasonalities. So a typical time-series kind of algorithm again, the context was a big feature, but it tries to basically understand how at each individual customer level is a behavior. Is he paying within time? Or is he paying early and claiming rebates? So, the system tries to analyze that, and then it takes the current open invoices and adds on top of that this algorithm. First to understand how these open invoices will be paid in the short term. And for the long term, the behavior that we have gathered over the last two or three years that is applied to come to a longer forecast, like the six months forecast. Because you will not have open invoices for six months per year. So that’s about the setup of how the data is fed into the system. And the system is calculating that. Now come into the variance analysis and the model setup. Now, once we start putting the models, again, this model runs on different variables, like your data could have a lot of variables. It could be seasonality, geography, a bunch of things. So the out of the box system takes this variable and even before we deploy, there are multiple models we create internally as well. And we try to compare for a week or a four week period. We try to compare what our model is forecasted and what is your actual data. Since you know, for these three or four week periods, we try to get even your actual data from your bank systems. And try to compare, let’s say for the case of AR, we try to compare the receipts that we thought you will be getting on a certain date versus your bank data to see what has actually happened. And then out of these three or four models once we choose, we decide that there’s a consistent pattern just done with a certain model, we finalize that model for you. And then we take it on from there. And even after that, the system starts tracking the comparison, which is where the variance analysis comes into the picture. And that’s where you start getting, seeing the data to see how the model is performing. Of course, if for some reason the model does not perform, or you see very high spikes or variances, and again, there are changes in models and all that. And the screen the Excel on the web that Amber was showing is basically where we put the entire model output into an Excel format, your typical Excel-based cash flow format. And if you have any additional information where you say, “Hey, you know what, I think this customer is not going to pay next week or the following week but will pay two weeks later,” have that information you can punch it in or there are certain cash flow categories which system is not focused at play taxes or payroll. Not a care system to forecast, because you very well know what the numbers are gonna be, you can punch them indirectly and the additional functionality you will have is the roadmap that we have. Multiple Model Management is what we call that. Like for example, you may say, you know what, I think this is what is going to be a slight change to create model one, you create another model and have this multiple sets of data, but pure executive management you say “Okay, this is what I think is gonna happen.” Your process for only one model, but the rest of the models you keep for your comparison, in the case in future you want to change it to another model or something like that. Okay, so that’s what happens in the system there. Any questions?

[18:52] Audience Speaker:

Hi, Natalie Simpson with Flowers Foods. You said currently, you only go out for six months and I think yesterday in one of the sessions, I saw the 2020 forecast for the Highradius Treasury solution. Is there a long term forecasting planned for 2020?

[19:13]Narender Nimmagadda:

So, there again multiple ways we can see this. One thing is what we have noticed and then based on the testing that we have done for the past one year when we are building a system is, the more you’re trying to see into the future, your forecast accuracy drops down. You may want to forecast, you may get a forecast, but you really can’t trust a forecast and just not much of us. And that’s where we thought, again, there are two separate sections that we see in the industry. So you have enough financial planning and analysis, where the forecasts are more generated by your own team, maybe the marketing team who probably have a better understanding of the market or maybe they’re planning something like you know, they were planning an ad campaign is something which the system will not be able to forecast. So that’s what we thought, mostly the treasury guys are doing month forecasts or maybe a quarter forecast. So we try to extend it to six months, which is what you want to do to manage your finances. And a one year forecast to use focus is not something we are focusing on at this point in time. But once the system gets better, it has more of your data. Let’s say we have five years of your historical data, maybe the system can do well. So there is no limitation from the system as such, what how much you can forecast, but the question of the accuracy comes into the picture. And that’s when we thought six months is an ideal state. Any other questions that I can hit?

[20:44] Narender Nimmagadda:

Okay, we have a treasury booth out there where we have much more detailed demos that are happening and we’ll be glad to answer any of your questions on one on one level, anything specific to your business or what you are doing or what we have in the roadmap, etc.

[21:13] Audience Speaker:

How does the system handle the forecasting cash flows in various currencies and then managing the variances at that level?

[21:27] Narender Nimmagadda:

Okay, let me repeat the question. So, you’re trying to understand how the forecasting handles the currency conversions and all that. So presently the system that we are working on is not into the currency and forecasting the currency patterns and looking into currency hedging. That’s not something that we are looking at. Typically the system does a forecast of a certain value and assumes the current exchange rate and it tries to show you a forecast. But once you have this and then let’s say if you have some other system where you are able to get currency forecasts then you can probably use that currency forecast in this Excel model, do your own math and come up with a different set of forecasts, which according to you will match the currency expectations. Like let’s say you have a business in Europe, and you’re expecting a 10% drop because of Brexit or whatever, I don’t know. So your system forecasts a certain value, and then you put a new currency rate and you try to get different forecast values. That’s what the power of excellent service is, right. So we are not into that, probably not our forte at this point, maybe some point we might get into that as well.

[22:39] Amber Thompson:

I think we are running a little bit low on time. So we’re gonna finish up here. And if there are any questions at the end that we can’t get into, you can always come to talk to us at the booth, and we’ll be able to answer those for you. So right now, our next steps are, of course, tuning that model and making it 90 plus percent accurate. And then we have two options that we discussed. One of them was to stick with forecasting in the US, and including those other cash flow categories like accounts payable, and other major items as well. Our other option was to expand into the entire region of North America. And with a discussion with the song, we were able to confirm that his biggest pressure point is AR. So we’re going with expanding our scope to more of a regional item. Then, in phase two, which will be coming shortly after that, we will add on those cash flow categories and accounts payable and assure that everyone in North America who needs access to the system do them viewers or contributors will have access on role-based access So that means that especially with the high sheets that we showed you, right? And they’ll be able to selectively view or edit those items as needed. This ultimately would bring us to our end case where we are forecasting, all kinds of work. So, now we have a few more minutes for any questions that you guys might have.

[24:24] Narender Nimmagadda:

Thank you. Have a good night.