GRA Partner Carter McNabb discusses Machine Learning (ML) and Artificial Intelligence (AI), and the significant opportunities for supply chains – specifically in relation to demand planning and inventory optimisation as well as sales, inventory and operations planning (SIOP).
In this presentation to the ASCI audience in Melbourne, Carter looks at what’s real and what’s hype in ML and AI in 2019. He reveals what systems and toolsets are utilising these cutting edge capabilities to help improve demand planning outcomes.
Following Carter's presenation is a demonstration of the GAINS advanced planning system by GRA Partner Luke Tomkin and GRA Manager Nathan Singhavong. The demonstration takes the audience through the three essential modules of the system: Demand Planning, Inventory Optimisation and Replenishment Planning and how ML and AI are being applied to great effect in these areas.
GAINS was recently rated as the leading Supply Chain Planning (SCP) toolset in Gartner Research’s 2018 Magic Quadrant report. For more information visit www.gains.net.au
If you would like to view just the GAINS demonstration watch it here:
If you would like to view just Carter’s presentation, watch it here:
Carter McNabb: Machine learning artificial intelligence, it's a very broad topic, very big field. I think Gartner says that the industry is very much still in the hype cycle. So the idea that when a new technology comes in, there's some kind of ear pleasing words that get thrown around, and there's this idea that this thing is now the answer.
Generally there tends to be a lot of hype, and a lot of excitement. And then, over time, the actual applications that work start to settle in, and it becomes part of the fabric of how we do things, as opposed to the answer.
What I want to do today is really focus on a very specific aspect of machine learning and artificial intelligence, which is its applications in demand planning. And I'll talk a little more about how we see demand planning, as well.
But again this is just a narrow and a very specific topic, because again machine learning in AIs is as broad as, in someways, the internet. It can apply to just about anything in the supply chain, depending on the use case, depending on the data availability, depending on the mathematical solving construct, and depending on whether or not it actually makes sense.
So we're going to narrow in on that a little bit. What is it? I mean I was contemplating opening with a discussion about what do we mean by intelligence, anyway, but that's a little too philosophical and broad. I love that stuff. But it might bore most of you very quickly. But what it is, is something that looks at large sets of data. And so, the benefit of it is it often can be disparate, not connected, and not similarly formatted sets of data.
And what machine learning does is will go out and look at various sets of data, and tries to build, I guess what we'd call, the relationship between correlation and causation. In other words, when this happens over here, does that create a relationship to something over here? And not only that, but is there a kind of a lead or lag effec? Because sometimes something happens here, but then there's not an implication, or an impact, until, say, six months later.
So there's kind of a lead and lag with these sorts of things, as well. What it's trying to do is look for these correlations, look for correlation and causation. And then, in the context of at least of demand and planning is then to be able to make automatic suggestions back, in terms of how to influence the forecast moving forward.
What we're going to do today is we're going to just give you a fairly finite example of what some of this looks like, and in forecasting, and inventory optimisation, and in replenishment planning.
The way we define demand planning, is that, to plan demand, it requires all of those elements. We have to have a forecast. We have to have an optimised inventory profile. And we also have to have a replenishment plan, to make sure we have the right stock, in the right quantity, right location, right cost, and so on. All those are important.
The tools that we're going to show you today, at least in the US, is integrated with what's called the FRED, which is the Federal Reserve Economic Database. And what it does is it, for example, it generates a baseline forecast. It also looks at history. But what it's doing is going into the Federal Reserve Economic Database, and saying things like, well, given interest rates, and fiscal and monetary policy, housing starts, and so on, as we see those, that data, change, does that appear to have any sort of influence on the historical demand? And then, based on that, if we have a sense of what's going to happen in the future, how might we actually alter the forecast, moving forward?
That's one very broad example. What it's doing, it's literally just going to the Federal Reserve Economic Database, looking at all the relationships between the data, and then also looking at what's happening in the history of the products that this business is actually supplying, and then making adjustments on that basis.
What do I mean by big data? Again, this isn't, really not a definitive answer. This is, again, more in relation to the demand planning aspect we're talking about today. But I just mention economic databases. So that may or may not have an implication.
I know one of the companies that uses a particular technology we'll talk about today, one of the companies that used it in the US was looking at the volume of steel imports, because, typically, there was a lower price. So what they would see, if the volume of steel imports increased, the demand for their products would decrease. And it was usually four to six months later. They worked out that correlation, were able to factor that into their sales and operations planning, and the demand planning, as well.
Weather's always a big one, particularly in retail. Or maybe if we're selling a very seasonal product, like ice cream, or something like that. Weather can have an impact. And, look, most of the cases I'm thinking about today, are really more around traditional retail, distribution, manufacturing, service and repair type businesses. Obviously weather and big data, I've seen some great applications for that in agriculture. That has a very clear implication.
Now, the interesting thing about some of that stuff, though, is I saw a great case study on an organisation that's based out of New York that grows food, basically, inside a warehouse. It's all temperature controlled. And the light is regulated, as well, because it's artificial light. And the interesting thing is they use machine learning to work out which inputs, at what times, effect crop yield.
Now, in that case, what I would call the relationship between causation and correlation, is very clear. There's a correlation, and you kind of know what's causing it.
Now, the interesting thing, one of the questions that comes out of this, is, just 'cause I'm seeing a correlation between something, did it cause it? Does that make sense? In other words, I notice that most people here are wearing darkly coloured jackets. It seems like, every time we have these breakfasts, darkly coloured jackets appear. Does the breakfast cause the darkly coloured jacket, or is it something else?
And so that's one of the questions to look at, as well. I might see a correlation. Does it mean it actually is the reason? Perhaps, perhaps not. Okay. We've got things like weather, temperature, price and cost changes.
Again, if we're changing the selling price of our products, if our competitors are changing their selling prices. Where I think is a big opportunity in machine learning and AI, is what I call the horizontal relationships in demand planning. Because oftentimes the business will say, well, if this product's increasing, it generally means that these go down.
What does that do for the overall category? We need to have a look at that, as well. In a big data sense, if you're looking at historical demand, and you're saying, here are all my products. And I'm kind of noticing that, as certain ones go up, other ones go down. Well we can start automatically building that in, versus just looking at a singular line-by-line forecast. Because the standard approach, generally, is to take the history and project forward, oversimplifying a bit, line by line.
With machine learning, we now can start taking a broader perspective in saying, well, what's actually happening in the population of the things that we supply and sell, when that sort of stuff occurs? Price and cost changes. They're going competitor activity, as well. If you're in a very price-sensitive environment that's heavily promotions-driven, and again we have a sense, both for what our promotions have been, and what our competitor activity is, as well.
Again, it can make some assessments as to whether or not that's going to have an implication, and you can do some scenario planning, and demand planning, around that, et cetera. Interest rates. I'll link that back to the economic database type of application.
Point of sale data. What's actually happening at the most granular level in terms of transactions, and can we make any relationships, correlations, between what's happening with our point of sale data, and that giving us an indication that, for example, well, if the point of sale data's a lot higher in the morning, perhaps there's not as much in the afternoon.
And again, if we have short shelf-life product, and frequent deliveries, that can really change the granularity and the accuracy of what we do in the supply chain as well. Population growth. What's actually happening with population in different geographies. And so it might not even be growth. It might be stagnation or decline, as well, relative to other parts.
If we're seeing that sort of information, and seeing that into the future, what might that tell us about where to invest, what sort of supply chain resources we want to have in place and what the delivery offer might look like, as well?
Customer order history. We'll talk about this a bit later, and show you some things in the actual demonstration. But it's one thing to say, well, this is what our demand was in a given period. Because, often with the forecasting, demand planning, we're working off monthly buckets. We might be working off weekly buckets. We might be working off daily buckets. But if we start looking at the actual customers and customer order profiles, and the lumpiness of that, if we start to get a clear picture around that, that can also start to help get more specific around our safety stock policies, as well.
And, actually, when we order and where we meet the customer. Because, again, I was out at a business yesterday and had this conversation. And they were using averaging techniques. The idea was, and they were a seasonal business, so, as the season starts to ramp up, on average, don't have enough. Then it peaks and starts to go down.
And so, in other words, we get the total amount right over the peak. It's just that we have half as much as we need at the start, and twice as much as we need at the end. So, not really a slam dunk. But, on average, it's pretty good.
The idea is that, if we get more and more granular, we can get that supply chain more synchronised and aligned to that.
Another one's search engine traffic. I'm seeing a lot of that with the growth of online. A lot of the retail businesses we work with have both an online and a physical presence. And what they're saying is we can start to get a sense for what's going to happen in the store with online orders, by the types of searches we're seeing, the volume of searches we're seeing, and so on. So that can now become part of an automated process in your demand planning, as well.
These are just some. It's by no means exhaustive. But hopefully that gives you flavours for some of the types of data sets that can be incorporated.
The other thing to keep in mind, too, and we'll talk about some applications, is that, depending on how big this gets, there's some serious, pretty serious, computing processing, going on here. The question is, is it worth it? Sometimes I think doing some sort of a proof of concept or building those relationships and correlations. Actually seeing that it does give you some insight is a really good way to go, before just kind of pushing the ship in the water, yay.
There's a few areas that AI machine learning. And by the way, just to be clear, and the pure mathematicians in the room will probably raise their eyebrows and look at me. But when we talk artificial intelligence, machine learning, and recurrent neural networks, largely the same thing. If I've got artificial intelligence, do I also need machine learning and the recurrent neural network?
Not necessarily. They largely do the same thing. There are some subtle differences that can be profound. But, in general, it's the same idea. Given what we talked about, in terms of the big data, there's another point around this, is that to use artificial intelligence and machine learning, and particularly in the tool set that we're going to show you today, you don't actually have to have any external data. This is the other point, is that you can work entirely with the data that you have in your own organisation.
One is promotions planning. I talked about using extrinsic, or external, information in promotions planning. That might look like, again, what our competitors are doing. But there's another piece, which is we planned to do this promotion, with these attributes, over this timeframe, and so on.
This is what we thought was going to happen as a result. But then this is what actually did happen. And not only that, after we ran the promotion, we did get an uplift. But then, after a while, we saw a big dip in sales, as well. Is that what we wanted to have happen?
So, again, one of the good things about AI and machine learning is that we don't necessarily have to go out and start boiling the ocean.
We can start just with some general information and get a lot of value out of that. To start getting a sense for, if we run the promotion at this time, in this way, what happens, and are there some other ways of doing that? And also, do we understand, again, that horizontal impact in the business?
Leading indicator analysis, you'll see an example of this today. This is one where we might do price changes, sales price changes. And again, what the technology does is basically determines if there is a statistical relationship, an actual correlation, between those things.
And then again, we'll recommend an automatic adjustment based on that. I think in the example, we're going to see it's not only an automatic adjustment, but it's also showing that, well, if we reduce the price of a 1.5 ml bottle of water closer to the 600 ml, some people switch out the 600 to the 1.5. And that's good to know.
Not just that it's going to create more demand, but there'll actually be a switch. Capacity and resource planning. This is where, if we've got a situation, and this is a very broad topic. It could be in a manufacturing context. It could be in a warehousing context. And this could be around seasonal peaks and troughs. It's one thing to look at warehouse capacity as the box all text. But a lot of times, capacity is impacted by the processing capacity, as well, and what time things arrive. What is our put away rate? And so on. What we can start looking at is, well, it's wonderful to basically have all this product coming in to meet the promotion.
But at the same time, if we've seen some issues with capacity in the past, is if you start to look at that and start to recommend a different, say, replenishment profile to suit the capacity in the warehouse, as well.
We've also seen some stuff, we've been doing some work recently in service supply chains. This is like our in-home care, and these sorts of things. Some of the applications we've seen are, it's the availability of qualified staff with certain skills, at a certain time, to do work in certain environments. And there was some demand planning around that. But again, the AI piece is looking at, well, can we put these people into the field? And they were assigned to do the work. What actually happened and what can we learn from what we're seeing about the actual behaviour, that whole process.
I think this is a really exciting one, is the new item introduction and NPD. The reason why I'm saying real exciting is because, I'll talk a little bit about the hype and reality. 'Cause there's just kind of like, we got to do AI, machine learning, go! But, well, I tend to do this, I just say what I need to say, and then we'll get to the slide.
But AI machine learning, it's a process of subtle refinement on something that exists. If you buy a block of land, and it's got trees all over it, and you want to build a nice little country house, or something, you're going to need a landscaper and a gardener. But you don't get the gardener to do the landscaping.
And what I mean by that is that, and I'll show you some evidence for this, as well, and I actually heard it, I presented on this topic, at the Australian Food and Grocery Council's Supply Chain Seminar. And both Johnson & Johnson and Coles said the same thing, is that we try to use AI machine learning for our base forecasting. Don't do that. Because it's too fine tuned. It's too variable, so it kind of gets a bit erratic. And there's quite a bit of university research that says this, as well, is that you need a very good strong base circle forecast, to anchor the MI to say, well, that's the forecast, and then the AI can do some refinement.
But it's not the base forecasting approach. And again, we're seeing that pretty clearly. Just again, it's not like we're here and we don't have this, so we'll just jump into AI and that'll solve our problems. It's not the play.
New item introduction, NPD, is a great one, because that's an area where a lot of the traditional forecasting and planning tools don't do as well. They give you the ability, in a lot of cases, to take history from other items. But you kind of set up that profile when you do the forecast.
This is now saying, again, looking at a broader set of attributes. Launch date, launch price, the amount we've made available, what discussions we've had with the customer, and so on. What actually happens, in terms of the uptake? And what are we seeing on an ongoing basis that's happening? And how do we use that to help, to allow the AI, the machine learning, to give us some information about what we can reasonable expect, as opposed to tell the machine what we expect, and then plan to it that way. Yeah, seems to be moving into exciting applications in that area.
Network inventory optimisation. Again, I mean, there's a whole bunch of stuff around route planning and supply chain optimisation. But what I mean by network optimisation is, it's not question of the network itself. But my question is, do I have a whole bunch of, do I consolidate all my stock and ship out? Or does it make sense to hold a bit here, and a little bit here, because, based on my lead time and demand variability, and cost, and, that makes more sense. Generally, what we find is businesses will make those stocking policy decisions and network flow decisions, and they become somewhat fixed.
They might move it every now and then. But the machine learning might look at it and say, well actually, it might make sense to consolidate the safety stock to a point. Or, because the behaviour of the item has changed, we now want to bring some more safety stock to this point, and then change what's held at the hub. This is that whole hub and spoke question. Do we go direct? Do we have multiple steps? That all comes down to things like lead time and cost. And so there's some pretty powerful applications in that, as well.
Automated replenishment and supply sourcing. The tool that we will show you today has the ability to actually look at. What it does is it learns from planners. In other words, as we get recommended orders, it'll build an order.
Say, well, here's the order, and I build it, and I try to fill, I put it in layers and I try to fill up the container, and I do some pull forward, that sort of thing. And I place the order. I've placed the order against the forecast. It's arrived. Do you know what actually happened? The system will actually look at all of those things and start to automate that. And again, there's pieces. In other words, we could say we're only comfortable for it to do that if the order value is less than X, and the forecast accuracy is quite good, and, and, and. And then we're happy for it to do that, or we can open it up a bit more broadly. But again, it learns from what it's actually seeing. Always a fun discussion in forecasting and planning, which is the master data management piece, which is, yeah that's what our history is, but we're out of stock.
One of the things that we're also seeing is AI machine learning being used to go back and make automated adjustments to the history to give us a better history profile to do our forward planning. In other words, if there was a shutdown, there was a supply shortage, something happened, as opposed to having to go back and fill in the blanks. This can do some of blank filling a bit more intelligently.
And some of the blank filling approaches that I've seen, that exist now, tend to be a bit crude in some cases. They just take a last point and the next point, and just draw a line between 'em. Maybe, maybe not. And again, we've talked about price changes, as well. Again there are many others.
This was just trying to give you just a bit of an eight course degustation menu for what this might look like. Where we're probably seeing the most interest at the moment, as opposed to the most actual benefit, but the most interest is, there've been the age old questions around how do we get more intelligence around that? What happens if we make price changes? And what about our promotional activity relative to our competitors? Surely that influences things? And that's kind of been a little bit of a subjective black art. And so there's definitely some interest in there.
But that also underscores the point to me, is that I think it also over emphasises the importance of forecasting. Forecasting is, I call it the holy trinity, in demand planning. You've got your forecasting, inventory optimisation, and replenishment planning. So far we're, with those points we're still only focusing on the forecasting, which is really helpful. But we need all of those elements to really deliver the service, and the cost, and the working capital result. What's available? I'm just, there's more than two blue boxes out there.
But what we've seen is Amazon Web Services has some machine learning forecasting. Basically, you can go on to Amazon, subscribe to it, throw some data out there, get a forecast back. Again, and I'll talk about the reality and the hype in just a minute. I'd say that that's interesting. I kind of like the idea. The issue I have is that you have to be very careful, again, about the use case and what you're expecting. Because if we use machine learning for our base forecasting, it doesn't give as good as results as a lot of the standard forecasting techniques we have. It's the refinement process. Just keep that in mind. It's a bit like using a toothbrush to clean your windows. there are better ways of doing that.
There are also some systems and integrated demand and supply chain systems that, or best-of-breed tool sets, that I think, then, if we add machine learning to that, and we build on it, we've got kind of the fundamental prerequisites already in place, and then we can take advantage of the next step. Thanks Nathan.
And some of the research, I actually found quite a bit of this stuff. But this was a university-based study, where basically they looked at what they consider traditional. And I actually looked up what they meant by traditional. Let me back up. This was basically the line, saying, let's take similar data sets and let's use a machine learning approach, and kind of the more traditional just using the intrinsic history of the item, the forecast forward. And actually, looking at what they call their traditional forecasting techniques, they were pretty basic forecasting techniques.
A lot of the tools we're familiar with have things that were far more grunty than what this was. But even so, the traditional techniques outperformed the ML techniques, sorry, the AI ML techniques, on a line by line basis. The opportunity comes to combine them. It's not a either, or. It's a and. Does that make sense? It's not a take one or the other. It's something, it's sort of the next step.
The latest thing is we heard that both J&J, Coles, told us that that was their experience. At least one of those businesses actually changed their base forecasting to use that, and they actually found that it was decreasing results, so they shifted back. Whereas, what Coles is saying is they're definitely still using their base forecasting approach, but they're actually using some very granular store level data. Weather matched with foot traffic, and things like that, to then use machine learning to feed back into the base forecast. I just want to make that point. It's an 'and', not an 'or'.
I keep talking about the holy trinity in demand planning. The core capability is the ability to generate a strong statistical forecast. And there's a whole universe of ways of doing that, as well. But, basically, we want something that actually looks at historical data, does some analysis and testing, and saying this is the most accurate forecast, and what my capability set is, if there is. I not only get that, I get to sit through the error in that forecast.
My inventory optimisation approach will take things like my service levels, my service level target, my costs, my constraints, and come up with a replenishment quantity or frequency, and a safety stock that's cost optimised and service level driven. And once I have my demand plan and my optimised inventory profile, then I can start doing my supply chain planning, replenishment planning, transportation planning and routing, really coming up with how do I come up with a set of orders that, again, allows me to meet my service levels at the lowest cost and with the least amount of effort.
So that's that fundamental capabilities stack. And this has historical data and has arrows going up, because, basically, these sorts of inputs, or these sorts of capabilities, do use historical data, and sometimes some future data, to get a picture of what needs to happen.
What we're also saying, though, is if we now take the big data approach, remember some of the things I put up there before. You can put them in economic databases, if you want to go that far. Price changes, weather changes, internet traffic, social media engines, and so on, and start to say, well, based on that, I'm now going to run the machine learning or AI layer on that, and then bring that information in.
So it's kind of a bottom-up and a top-down at the same time. And if you think about the old school process, it used to be what it looked like you would use the base history, and then they would use this intelligence to look at the data and say what changes. Now we've augmenting that intelligence with some additional mathematics to help with that process.
To be clear, it's certainly not a replacement for human intuition insight. But, again it gives us and insight. And you mentioned a great point about the insight you got, about it's actually consecutive number of days, not a monthly figure, which is really great. We can now bring that into our sales, inventory, and operations planning process, because it gives us the ability to model and make tradeoffs like we really haven't before.
So we've heard a lot about the need to have better scenario planning capabilities. I think this is where AI machine learning can really help, as well. We marry those things up. But what I also want to point out is that we've got, there's a line going from historical data to machine learning, as well.
Why is that? Because, in the tool set that we're about to show you, and again we're just going to give you just a small taste of some of these things, but what it actually does is it looks at what it forecast, versus what actually happened, versus what it was adjusted to, versus what actually happened. It's also looking at what it recommended in terms of safety stocks and order profiles, what the planning team actually did relative to that, and again what happened.
And so what it does is it starts to say, well, this is what the base of mathematics tell us. But now what we're seeing is actually happening, we can now make some, again, some finer tuned adjustments within that, because we're getting more and more granular data.
And so, again, it's kind of like your point about the number of days of certain temperature in terms of its influence on ice cream demand. It's also interesting to have the right amount of safety stock arrive on Thursday afternoon when I needed it on Thursday morning.
So again, we're getting more particular and more granular, which allows us to be more effective, both from a cost and service level perspective. So, just, I'm about to handover to my lovely colleague Luke, who's standing in the back.
But before I do that, we'll just go to the last slide. The tool set we're going to show you is one that we've been aware of for a long time. It's called GAINS. It does have that complete capabilities stack, and has the machine learning and AI built into it. I'll hand over to Luke to tell you a little bit more about that, and then Nathan will take us through a very curated and very short demo, just to give you a taste.
Luke Tomkin: As Carter was speaking about, machine learning's got wide and broad applicability to a lot of pieces, a lot of areas.
Obviously, supply chain professionals are really focused in certain aspects of that machine learning. What we're going to try to do now, is really just bring it down to a couple of real, concrete examples, to show you how it works in practise.
Carter's earlier slide showed three things that we would say is at the core fundamental of supply chain planning. We've seen quite distinct steps.
There's been a demand planning step. The demand planning step, what we're basically trying to do is work out what our customers are going to consume or buy. The second step is an inventory optimisation step. And what we're trying to do there is basically deliver to that customer what they want at the least total cost to our organisation, save as much money as we can, make as much profit as we can. And our third step is making sure that we can get that product where it needs to be, on time.
So if you think about demand planning, inventory optimisation, and replenishment planning. What we're going to show you is basically what's the foundation or that base that we need to build the MI off, for each of those three steps. If we think about an event planning system, which is what we're going to demonstrate, that system starts in the ERP system. The ERP system, as probably everyone's aware in this room, holds all of the transactional information, the master data.
And we have the advanced planning system do some more sophisticated work on top of that, so we can actually take intelligent action. So, if you take demand planning, for example, the foundation, or base level, of requirement that we would suggest that you need before you move into MI, would be the ability for the system to generate a statistical forecast, by itself, based off consumption or demand history.
Using a tournament forecast selection process, to be able to identify this is the forecast which will give us the best result with less error. Pretty straight forward? That's the base level that we need.
The second step in that is the inventory optimisation step. What we would say would be the base foundation of what you should be considering before moving to MI when you're talking about inventory optimisation is all about achieving a particular level of service to your customer, at least total cost.
What that system should be able to do is, say, for every product, every location within your network, here's the amount of safety stock we need to hold based off the cost of holding inventory, cost of moving inventory, and the level of error we have in our supply, our supply variability, and our demand variability.
And so, if you're a large retailer, for instance, that might be three million different stocking locations throughout your network, and that's the sort of capability that we would look for in a business. If they're after MI, we would try and establish that before moving into that MI space.
The third area is the replenishment planning. And the most base foundation expectation that we would have is that the system would bring up to you, by exception, on a daily basis, anything that you need to take action on, in terms of an overstock, understock, or an expedite.
Basically anything that's in a balanced inventory position, a user shouldn't have to look at it or consider it. That's what we're talking about the foundation, or the base that we're after. Now if we try and start overlaying MI, what are we looking to do?
Rather, in the demand planning space, rather than just relying on that consumption history that we have in our system, the transactions, we start to bring in what we would say extrinsic data, data external to the ERP system. And that external data could be, for instance, potentially, competitor pricing information. It could be the weather. And to get to that we need to actually, you know someone asked that question before, it's actually finding out what are those elements that do have good correlation, actually do have an impact. And they could be lagging, as Carter mentioned before. If something were to happen four months ago, it has an impact on us now. Like housing starts would often be an indication of that in the construction industry.
In the inventory optimisation space, rather than looking at what a particular product does in terms of how do we get the cost-optimal service for that particular part, that particular location, we can consider the entire network.
So rather than, potentially, having a 99% service level which delivers that service that will have very high safety stock at every location, we might be able to say, if we hold a 99.5% service level at one location, we can get away with holding a 60% service level somewhere else, and just cop the cost of an expedite to get to that other location in time, and thereby save money. If we start to think about these in more holistically, in moving from inventory optimisation to what we'd say is profit optimisation.
And then the third example, replenishment planning, rather than just showing up exceptions that which a user needs to review and approve, or modify, we can now get the machine to say, well, what would've the user done when these exceptions get brought up? And we can put on those. And the system itself can apply the logic to say, under this scenario, our user does not need to intervene. It will automatically know to adjust it, or also cancel it, or to modify it.
So, I might just hand it over to Nathan now to start showing exactly how this works. But when you think about this demonstration, it's in those three components. The demand planning component, inventory optimisation, and replenishment planning. There's just a taste of what MI can do. It's not the whole gamut.
Nathan Singhavong: I'm Nathan from GRA, a manager at GRA. Yeah, good intro . I've been involved with GRA, now, for about five years, and have been working primarily on GAINS implementation. I've done a few implementations with Wesfarmers Industrial Safety. I've done one for 7-Eleven, currently. Large scale implementations that have various different types of challenges.
This is the GAINS application. I'll walk you through some of the things that Carter and Luke have both touched on, looking at the three modules: demand planning, inventory optimisation, and replenishment planning.
But first I want to speak to that base level of forecasting that Luke and Carter had mentioned. An organisation is built up, not just on the number of SKUs, but the number of SKU location combinations. You might think that we're only ranging 1000 SKUs. But we've got 1000 SKUS across 1000 locations. That's a million data points that we need to forecast this.
Now, if you multiply that by the number of sales channels that you have. You might have an online sales channel. You might have a retail sales channel. That's two. That one million forecasting data points then becomes two million.
Now, from a user perspective, you're going to need a massive team of hundreds and hundreds of people just to review each and every single line item.
So what GAINS does is it boils down that two million data points into what you really need to focus on. This is an entry page onto GAINS where we actually have a series of exceptions. We've broken it down into demand planning exceptions, both at a store level, DC level, if you wish.
You can also look at specific business processes, such as event management, store management, all those kind of things. If promotional planning is key to the business, we can create certain exceptions, as well, to highlight those. There's also the replenishment recommendations and exceptions.
We can walk through what is the base level type of exception that you might see on a replenishment. And also we'll walk through what machine learning will then do, as well, to minimise that data set that you actually need to look at day to day.
These little boxes are kind of like a summary of the workflow that a planner needs to do during their review period. We can set up a review period that looks at a monthly review period, or a weekly review period. In this particular data set, I think there's about 13,000 item location combinations.
You'll see that there won't be 13,000 recommendations that you need to review. But it's really looking at specific examples and then say, within this review period, you've actually got 437 forecasts that are either biased or volatile.
Now, during the review period, we've actually done some forecast approval. The 's actually looked at it. We've actually moved that down to 336. For the rest of the period, we actually have 336 forecast reviews to make.
What's nice about this view, is that it not only gives the users a sense of where they're at during the week, or during the month, but it also allows planning team members to actually analyse the process flow, as well, to go, okay well, this point is actually, even though this planner has 1,000 SKUs, this person's only got 50.
This one might actually be more highly volatile, more errors, so their work weight is actually much, much higher. It also allows a little bit of rebalancing, as well, between team members.
This screen is the base forecasting screen. It's what we call the Forecast Summary screen. There's multiple levels to this. You can either view two years of history, two years of forecasts, both at a monthly level, weekly level, and a daily level. We also have the ability to group by certain article attributes.
So, for example, if you wanted to review what a state was doing, you could roll it up at a state level. If you wanted to see specific sites, specific categories, inventory classes, it's all made pretty possible, and pretty easy to view everything at the aggregate level. The bottom level, here, is your lower level data. In this instance, it's basically an article at one particular location.
If you had article, by location, by sales channel, that would be displayed down at that lower level, as well. The row that's highlighted at the moment is the pastry level. And these are all the article locations that make up the pastry. There's beef pies, there's halal pies, sausage rolls. At a base level, GAINS will use intrinsic factors, such as sales history. We take up to three years of demand history to build up a history profile, which can be reviewed by a demand planner. It will then generate a statistical two year forecast.
And I'm just showing you this particular screen which runs through the various forecasting models that GAINS has. GAINS has, within itself, about 35 different forecasting models. If you think about those two million item, location, channel combinations, that's actually being multiplied by 35. Two million by 35 is 70. It's actually doing 70 forecast tournament selections every single night, or every single run. The amount of power behind this tool is quite significant. Each module, each batch run, it'll actually run through model one through to 35. It'll then rank each of these models in how they perform.
Carter mentioned that, out of the demand planning module, not only do you get a forecast, you actually get a forecast error. And GAINS has some innate ability to decide whether or not it matches the demand pattern. It'll look at it and say, that roughly matches our history; that's okay. It'll also flag certain forecasts for review if it's deemed to be biased or volatile, as well.
GAINS will automatically select the forecast model that has the least amount of forecast error, but also if that forecast model is biased, it'll say, maybe let's look at the next one down.
We don't want to choose a biased forecast as our main forecast, moving forward. The classic example of that is when you have new product launches. When you have a new product, history looks like that. Base model would say, forecast also looks like that. But at some point it's going to plateau. GAINS will say, this the best matching forecast, statistically. But, realistically, that's not the one that we should be applying. You need to look at it. What I will do now is I will show you what the machine learning layer is, on top of this, as well. When you would talk about machine learning, we can feed in extrinsic variables. Thing like weather, things like price changes, things like promotional activity.
Typically, when we go into an organisation they'll say, "it'll be great to know "what our competitors are doing. "If we can fit all that information in, "we'll be able to know what levers, "what they're doing, so that we can react to it."
The next question is how do we get that data. They go, "oh, that's actually really hard." Typically what we will suggest is you look at what internal information you have. Not just what's external, but things like price changes, price changes relative to other items, as well. The example that we're going to walk through is a 1.5 litre bottle of water.
This particular organisation has various sizes. They've got a 1.5 litre, two litre, 600 ml bottle. And what we're doing there, is we're mapping the actual retail price of each of them. But not just the retail price of the particular SKU, the retail price of the SKU, relative to the next size up or the next size down.
There is a correlation. If the price gap between the two bottles, the 600 ml bottle and the 1.5 litre bottle, shrinks, you'll find that some people actually go, it's actually a better value to buy the 1.5 litre bottle than the 600 ml bottle.
So you'll see that sales on the 600 ml bottle might decline. Sales of the 1.5 litre bottle might increase. And then how does that impact the two litre bottle, as well. I'll show you a quick example of that. This screen is what we call the Forecast Detail screen.
It's just a different view on the forecast, primarily at the monthly level. But what it provides us the ability to do is to analyse, not just what history was doing to it, but also what our extrinsic variables are doing to it, as well.
There are three data feeds that we are feeding into this. We're feeding in the retail price of the SKU. We're feeding in the price gap of this versus the 600 ml bottle. And we're also feeding in the percentage of sales of the category, as well. What was the percentage of sales within the month of this particular bottle of water versus the whole drinks category?
You might have taken 70% of the population of the sales, but it would learn from that and go, okay, what's the next forecast. If we're heading towards a warmer month, such as spring or summer, you would expect that the percentage of sales of water would increase, and decrease in soda.
And GAINS would be able to recommend that. GAINS also has the ability to look at leading indicators. Leading periods. Just because you make a change to the system, doesn't mean that it's going to impact it straightaway. If you make a change to the system, you might see it a month later, two months later, or three months later. GAINS will then play out each of these scenarios, and not just each of these variables. Your price gap now. It'll say what is the impact of the price gap if you change it now, versus a month's time, versus two months' time, versus three months' time. On this screen, these are the graphs that are being displayed. The green line displays your history profile across the last two years. The blue line displays the percentage of drink sales for this population.
We also have the retail price and the price gap. You will notice that, in this section here, we've got that check box there to say we're excluding insignificant variables. Out of all the data elements that we'd be feeding in, we've had one, two, three, four, variables, multiplied by six leading periods. That's 24 potential variables that it's looking at. It said that, for this particular item location, there are only two that actually have a significant impact on the history in the forecast. Out of those 24 variables, or 24 combinations, there are only two that matter to this particular item location combination. You might find that this bottle of water in Queensland will behave differently to Victoria, and there'll be different levers there, as well, just because of the sales pattern at the store level. One of the examples that we like to use there is gum boots. You might have gum boots that are stocked in Victoria, gum boots that are stocked in Queensland. In Queensland, you'll see that during the wetter months there might be some unexpected demand that pops up, because of flooding, et cetera, or that kind of stuff. If those events can be built into the system, GAINS has the ability to recognise and say there's a potential that this event might occur again in the future, so we need to uplift our demand profile.
As an input, these are the various price gap changes. You can see, relative to each month, what the retail price change was. You don't actually have to look at that information, it's there for those statistical minded people. What we also do is we also feed in future leading indicators. Not just what the retail price was two years ago, 23 months ago, 20 months ago. We also say what do we plan on putting in as a price increase? The reason why we say focus on what internal data you have first, is that it's a lot easier to control your own price, versus what your competitor's doing.
In this instance, we've fed in that, over the next two months, our retail price will be growing from $4.20, to $4.30, to $4.40. And there is also a price point difference, as well, between that and the 600 ml bottle.
What GAINS has then done to the data is it's looked at it and said, next month I was originally planning to sell 176, but because of my price gap difference, I'm now forecasting to sell 119 units. GAINS will do that automatic adjustment.
The reason why we limit it to these two, or three, potentially, data points is we don't want to touch the forecast too far ahead. We only want to look at a horizon that makes sense. So with this particular item, there's a lead time of a day to a week.
You don't want to be making changes to the forecast three months out, unless you'd have a known promotion. We've walked through the base forecasting models. GAINS will do a tournament selection based on 35 different algorithms. A user then can then overlay their own intelligence over the top of the history or forecasting profile, and the machine learning can then be able to enhance the process, as well. Following on from the forecasting module.
I'll take you to the inventory optimisation module and just walk through what that looks like, as well. When Luke mentioned earlier that we take into account a whole bunch of costs, we take into account the cost of holding stock, the cost of receipting stock, the cost of utilities, as well, to run your DCs, all those elements need to be taken into account.
What GAINS does is it looks at auto quantities, and says, okay, what is the most optimal auto quantity for this particular combination, for this particular item location combination? If you order in lots of ones, you might find yourself ordering every single day. If you order in lots of hundreds, you might find yourself ordering once a week.
GAINS will find that right balance between your ordering quantity and your safety stock. In this particular example, we've got a, I think it's a two day lead time. We have a carrying cost set to 15%. That basically says that for every dollar, I'm actually carrying 15 cents additional, over the next year. It'll cost us $13.00 for PO to get this into the network. And we also have the lead time variability.
What GAINS also does, at a base level, is it builds in your supply constraints. If the supplier says you can only order in lots of 20, GAINS needs to factor that into account as well. If GAINS says your optimal ordering quantity is 21, does that have to then round down to 20, or round up to 40? If it rounds up to 40, your safety stock will decrease. If you round, what is it, down to 20?
Your safety stock might have an increase, as well. There's a balancing act between these two parameters. The next level on top of that, is to do some what-if analysis. What if we had to expedite? How much additional cost would we incur if we had to expedite?
This is a graph that models your total annual cost versus your service level target. In this particular instance, I think we've set a service level target of 98%. At 98% our total annual cost is $650.00. If we were to drop our service level down to 70%, meaning that 30% of the time our customer walks into our store that will have, they'll have an empty shelf, that actual annual cost is about $500.00. If we were to do some modelling with, say, the normal cost of getting a PO in is $13.19. If we were to expedite it, we would mark that up to $20.00, because we need it urgently. It's out of cycle with our supplier. They say, "if you want it, "you have to pay a little bit more for it." What we've done here is we've entered in an expedite cost of $20.00. It's basically now said, well, for every single miss that you have, if you want to make the sale, you need to expedite it.
This has then added in an expedite cost component to the graph. And what this graph shows me is that it's actually more cost effective, over the year, to hold a higher level of service than to hold a lower level of service.
If I were to promise my customers 70% service level but I would actually have to incur additional expedite costs to make up those lost sales, I'm actually spending just under $900.00. Whereas, if I had, still, my 98-99% service level, I'm still only just incurring the $800.00 mark. There's a little bit of modelling that comes into this.
Where we find this really handy is when you're changing supply models. If you're moving from a local supplier to an import supplier, your lead times increase, but your cost might decrease, as well. What you can also build into here is the disservice cost. What happens if we lose a sale? Would we lose our customer forever? Will they go elsewhere, as well?
There are all these little factors that can play into this kind of model, not just to perform a little bit of what is GAINS doing, but also, what can that do in my new supply chain network, as well?
The next screen that I would like to show you is what we call the Autobuilder screen. That provides the users with exceptions to say, hey look, I think that these ones are going to be out of stock during our lead time, or projected to be out of stock in our lead time. And then it would flag for replenishment planners an action.
As Luke mentioned, we don't really care about our balance stock. There's no action to do there. There's no action to review. There will be no action. What I've clicked on is replenishment for today.
What GAINS can build in is also a dedicated planning cycle. So if you have a supplier that supplies into a particular store, Mondays and Fridays only, you'll only get recommendations on Mondays and Fridays.
What GAINS can also do is then throw up other exceptions to say, even though you're ordering Mondays and Fridays, you're actually projected to be out of stock on a Wednesday.
You might be actually incurring some additional expediting costs, as well, if you were to bring that in. This is what we call the Autobuilder screen. Top section, there, displays a few graphs. A lot of people like working with graphs.
A lot of people like working with data. Depends on the type of person you are. If I look at a graph, I can understand it instantly. If I look at data it takes me a little bit of time just to go what is this actually telling me? Some of the inputs that we can actually feed in are vendor minimums and vendor maximums.
Obviously, if you haven't hit your vendor minimum yet, you want GAINS to actually tell you, hey look, we haven't hit our minimum yet, what do we need to do to get that up to our maximum, sorry, to meet our minimums.
On the flip side, if we are over our approval limit. Let's say that you need approval from the supply chain manager if you're exceeding X amount of dollars. You can actually get GAINS to pull that back a little bit, as well, to say, we've got a limit of $1000.00. You're actually purchasing $2000.00 in there. If we drop it down to $1000.00, what, actually, will we be purchasing.
Rather than leaving the decision up to the demand plan, the replenishment planner, to just decide, hey look, I'm going to drip this one down, this one down, this one down. GAINS will actually say, you need these ones first, before you need the rest.
There's a summary section which shows you the post-order equivalent. In this one, we're actually shipping from the supplier 100351, into the receiving site 1003. This has been made up of four order lines, as well. So you can see here that we are ordering $500.00 worth. And we are well within our bounds. If we were to click on this particular item line, you can see there that there's actually only one order line being recommended for this particular store, that it comes up to $45.00.
Our minimum, actually, is $100.00. We need to do something about that, as well. GAINS does have the ability to use a set target quantity, where we can actually bring up the order value. When we use the target quantity, we can also bring it up, not just based on dollar amounts, we can do palletisation. Number of pallets, number of cartons, number of layers, as well as number of units. The example that Carter and I were at yesterday was we were at a musical organisation. And they've got a wall that they fill with guitars.
They actually have rounding factors, not based on volume or anything, but based on inches. And they say, "look, we don't have any storage space "in the back. "On the wall we can fit 12 guitars." "We can also feed into the fact "that GAINS says we've got eight on hand." GAINS might recommend another eight to be ordered. But we can only have space for four, so which four do we actually choose?
You can get GAINS to make that decision for you. Once GAINS has been up and running for a period of time, we can also establish some learning attributes as to what a replenishment planner has been doing. We can say that GAINS has recognised that for low cost, low lead time, low error items, typically the planner just goes accept the order. Classic example of that is rubber bands, right. Rubber bands are cheap, they come in small boxes. You might not think about them too much. You just go, the come in with an order, I'll just accept it. Whereas, if it's a bathtub that's big and bulky, and takes up lots of storage space within your DCs, GAINS might say, hey look, I'm not going to auto approve this one, because, typically, when our planner looks at it, they go, "yeah, I don't need 20 units of that, "takes too much space. "I'm going to drop it down to eight or 10." GAINS will know, I'm going to stay away from those ones. I'll look at these ones.
When we switch on this learning capability, GAINS has the ability to recognise certain patterns. In this particular data set, this one is actually an approval of an auto line, or a particular order. It looks at anything that is less than two days of lead time. The average cost is found to be, well, if it's under $6.60, generally the planners accept it, outright. If there's a lead time forecast error less than two units, as well, it'll say auto approve, auto approve.
This takes away a lot of the decision making that the planners have to do. If you think about what a replenishment planner has to do, day to day, they have to review order lines, accept, accept, accept. This can actually take away some of those mundane tasks, and actually focus them more on those containerisations, those big ticket items that are big and bulkies, rather than looking at the small ones that just run through the system, always.
That kind of rounds the GAINS demonstration. We looked at demand forecasting, which looks at building a base forecast based on your intrinsic data. Three years' worth of history. That then gets overlaid with user adjustments. We then overlay the machine learning aspect of it where we fit in extrinsic data. That can be pricing, weather, promotions, competitor, competitor's promotions, as well, competitor pricing. We then looked at inventory optimisation, which builds a safety stock and order quantity profile based on ordering costs, carrying costs.
You can then do a little bit of modelling on top of that, as well, to build on your expedite costs, disservice costs. What happens if I change my lead time? What happens if my cost increases over time? That can actually inform you how to build your future network. And on the replenishment side, it's really about taking away workload from replenishment planners, to focus on the big ticket items.
Carter McNabb: All in all, it's a really exciting area, I think, primarily, because it's now bringing into the realm of possibility the discussions that I know I've been having for the last 20 years about all this extra stuff that really didn't have a great way of being dealt with.
Again, we had some very powerful tools for forecasting and planning, but there's always this question about, but we know things happen when we make adjustments, or we change prices, or we input promotions.
But we're not really exactly sure how to kind of deal with that in a holistic way that gives us the actual insight to then go back and apply. It was kind of, there was kind of, the white art, if you will, and there was the black art, which was how do you deal with this sort of stuff. Now we have a way of dealing with it, which is great.
And so, again, the question is, I think is, does it make a difference? Nathan showed a really good example of a whole bunch of things that might have made a difference. But we got to see what actually does make a difference.
I think a really great way of moving into machine learning and artificial intelligence is some proof of concept. Is it worth the investment? Is it worth the computing power? Does it make the forecast and the plan more accurate? Let's have a look at that before, as I said, we go out and roll that out, because it is a process of refinement.
In a way, I think it's, the great news is it's actually here and it's working. I think the next five to seven years we'll see some really exciting developments in that area.
Reproduction of GRA whitepapers and articles
GRA permit the reproduction of GRA authored whitepapers and articles so long as all the following conditions are understood and met:
- Entire credit details must be included:
- Author's name(s)
- GRA name and contact details
- GRA URL link to the original article
- All hyperlinks within the article must also be retained
- Articles must not be resold
- GRA retain full copyright.
If you have any queries about reproducing a GRA article or whitepaper, please contact GRA Marketing