Lean and Forecasting
Dear Gemba Coach,
I’m in charge of the forecasting department in my company. My colleagues in production have been doing lean for several years and complain about my forecasts, particularly when it comes to what they call leveling. I have come across your column on lean.org and wondered if you could help out with some advice on how to forecast in a lean way?
Wow. Interesting and difficult question. There are a number of angles to this question, so let me attempt to clarify the topic somewhat. First, there are many different beliefs among lean practitioners about the reliability of forecasts (contrary to common opinion, experienced lean practitioners do not believe that all forecasts are wrong). Secondly, this issue translates into determining how to turn forecasts into leveled production plans and who owns or manages the inventory of finished products to absorb customer demand variations. Thirdly, there is the simple challenge of learning how to improve the exercise of doing forecasts themselves.
Please bear with me for the lengthy response to your question—but I must point out that there are many facets that are covered here. I guess the main point is that, in lean, surprising as it can be, we don’t try to forecast accurately. We try to have a robust ballpark idea of where the total volume of the market will be and then adapt our production processes to be flexible for last minute mix changes. Overall volume is about capacity planning (which doesn’t need to be detailed, but has a huge impact in terms of capital costs) whereas day-to-day mix requires production flexibility. Robust ballpark figures enable us to calculate takt times (which can’t always be changed easily), but on the other hand detailed day-to-day forecasts don’t help as we’ll try to follow real-time customer demand. This is the main change of emphasis for forecasters: better overall volume forecasts, and less detailed predictions.
In this sense, lean thinking assumes that volume can be forecasted quite reliably, but not mix. This means, for example, that from one year to the next, we can make reasonable assumptions about how many people will buy coats for winter in any given town. We can have high-level models of what percentage of the population will renew their coat. If we know people buy a coat every two years, this means that 50% of the population will buy a coat this year (with cars, for instance, the market is estimated at 10% of the total automotive park a year). Such estimates can be easily tested by traditional forecasting methods from historical data. The assumption is reasonable because it reflects an underlying process: so many people will purchase so many coats.
Predicting the Future
On the other hand, what is extremely hard to predict is what kind of coat will they buy: will the fashion be dark or light? Short or long? From historical data, we can tell that people will buy a certain amount of short coats every year, and a certain amount of long ones, but there remains a sizable portion of coats where long/short is decided on the spot according to the vagaries of fashion. This is very hard to forecast, unless you know someone with a good crystal ball.
The other thing you need the crystal ball for is knowing exactly when it’s going to get cold and trigger the coat-purchasing action. If cold hits suddenly in early fall, and you don’t have the coats available for sale you’re going to lose sales you’ll never recapture for the entire season (and which will lead you to unreasonable rebates later on to try and flog the stuff you’ve finally brought online). Conversely, if you get into the season with a huge pile of coats and cold days come late, fashion might have moved in an unpredictable way and you find yourself with a pile of coats people won’t buy (again, unless you discount them). Not fun.
So the lean assumption is that total volume is fairly predictable (worst case, predict what sales volume was last year), which makes sense because of the law of large numbers (the results of large samples are more trustworthy than smaller samples, because large samples have less variations than small.) Yet specific demand for options or models is very hard to predict from one day to the next. As a result we can separate forecast in an overall envelope, and then project the lowest quantity purchased for each option or model, and accept that there is an unpredictable part as a percentage of the whole. Not comfortable if you are a forecaster, but fact of life.
As a result, one can identify a few things that a lean forecaster should be very good at. First, tracking overall volume and producing as accurate as possible estimates of minimum, maximum and average expected total monthly volume. This is the basis of capacity estimates for the coming period. The second is to pick up and respond to rapid signals from the market to make educated guesses about the mix. The best forecast for the weather remains to predict yesterday’s weather and minding the signs for a sea-change. The forecaster needs to learn to identify the proper signs of immediate change and try to make short term guesses, not long term.
Now to scheduling. The big, big lean change is that we learn to see that most variation is created demand, as opposed to real market demand. MRP systems start scheduling by looking at customer orders, seeing what is in inventory, checking the level of safety stock and then producing a work request by the magic formula of CUSTOMER ORDERS – inventory + safety stock. Then it looks at batch sizes (in the system), capacity (in the system), and then produces a work order which gets sent to the shop floor, where it is then adapted to the issues of the moment. The magic formula turns out to be a great multiplier of variation. Suppose the customer is not ordering so much today, the inventory increases from yesterday’s production, so the calculated demand lowers, which is sound enough, except that the effect is not felt right away because of the batch sizes and the queue. Now, tomorrow, the customer orders double: the average, plus what he didn’t order yesterday. This creates a peak in calculated demand, just as production was reducing the production of these parts and so on. Daily pain.
The biggest difference with the lean model is that the planner will own the finished product stock. The planner now has to make trading decisions. The idea is to propose a production plan where the quantities of As, Bs and Cs to pick up from one line are the same every day for each day of one week. This means that the finished goods stock of As, Bs and Cs absorbs the natural variation of customer demand and buffers this information towards production: you stop adjusting production to every variation of your inventory. Essentially, as long as the inventory doesn’t hit the MAX alarm level or the MIN panic button, it’s left free to vary randomly. The trick lies in estimating correctly the correct MAX – MIN bracket to hold in stock, and this can be adjusted from one week to the next.
The question now becomes very different: it’s no longer about how many parts do I need to replenish, but how quickly will I get parts to replenish. The scope of inventory variation depends both on customer real variation, but also on replenishment lead-time, which means reducing batches. With smaller batches on the same cell, I can have a higher safety cushion with a lower overall inventory. This is classic lean stuff, but the trouble is that customers use MRPs as well, so many customer order fluctuation are also created demand. More pain.
So the third thing a lean planner has to be really good at is creating a stable, level weekly plan where it warns production that it’s going to pick up the same amount of As, Bs and Cs per day for every day of the coming week (barring holidays, shorter hours on some days and all other stuff one learn to work into the plan). In essence, it allows production to have a takt time by averaging the customer demand. Whether production chooses to make this in huge batches and hold huge stocks or in smaller batcher and have lower stocks in production’s problem, but what it doesn’t mean is that you need to physically distinguish inventory owned by logistics to cover customer demand from inventory owned by production (and kept in production area) corresponding to the batch size. With the pull system, planning is telling how much it will pick up in the day, and now logistics will come and withdraw a few boxes of As, Bs and Cs every half-hour: fractioning and mix, but this is a production issue, not a planning one.
The surprising upshot of this mental shift is that you will need less planners in terms of brute force, but will become desperate for a really good “master scheduler” to create the volume forecast and the short term leveled production plan. This is more a matter of skill and judgment than massive calculation and sophisticated software.
Don’t Forecast, Understand
None of this is easy as it requires a few serious mental makeovers. As always in lean, the key to these issues is to deeply understand the problems we’re trying to solve:
- First, we’re not so much trying to forecast as to understand our market. We need a model of the overall market consumption (what are the drivers) and so we need to get better and better at understand what moves a market and drives seasonality, such as it is.
- Secondly, we need to make short-range guesses in terms of which specific product customers will want now! This means working hard at improving the early warning system and having real-time information about what customers are doing right now. For instance, if your customer is also a business, its own production plan is not a state secret. Rather than forecasting on the historical data from customer orders through its MRP, you’ll better use your time by getting to know your opposite number who schedules the customer’s work plan and understand what they aim to build in the coming week. If you’re working B to C, then, same thing, rather than focusing on what’s held in the warehouses, you need to get real-time point of sale information to figure out what customer preferences are right now.
By understanding the overall volume and the immediate mix trend, you can then figure out the max/min tunnel for your key products and use your finished good inventory as a “shop” – improving your trading decisions of how much variations (hence inventory levels) you allow for, and when you hit the panic button (as in stop building more, or build some more immediately). You can measure your skill as a planner at (1) On-time-delivery, (2) inventory level and (3) the number of times you ask production to make an extra effort to save the day.
- Thirdly, you need to constantly improve your ability to give production a leveled schedule (the same quantities every day) and not panic on Thursdays and yell “forget the plan, I need more Cs”) – this ability is of course linked to the two previous points, but, per your question, your production colleagues will LOVE you if they now can work with stable schedules.
Well, love you, maybe not. The whole thing hangs, of course, on production’s ability to produce in small batches – to improve its flexibility, so that it can react swiftly to mix changes and solve your inventory problems (and theirs). Which often comes down to … relationships.
- There is a fourth skill you need to improve as a forecaster and planner, which is the quality of your downstream relationships with your internal customers and your upstream relationships with your internal suppliers. To lean the process, you need to get better, fresher information from the former and more flexibility from the latter. This is not a small challenge as they often won’t understand what you’re after.
Most conflicts, I find are born of misunderstandings of intentions – there are two aspects to your question: sure the technical side of “lean” planning and scheduling, but also, the realization that forecasting, planning, and scheduling are essentially teamwork jobs and that we need to repeatedly discuss issues across functions, in order to share a common understanding of (1) what we intend to do and why, and (2) the practical barriers that always come in the way. In most of the firms I can think of that have been successful with lean, the planner (master scheduler in our parlance) often becomes one of the most important people in the firm: he or she is the helmsman of the ship!
How Using Kanban Builds Trust
Kanban functions as a trust machine because everyone using it must understand what they have to do and why, says Michael Balle: "Our purpose here is to share our ideas on what we believe is important in lean thinking."
The Sanity of Just-in-Time
Path dependence is the worst enemy of smart resolution, argue the authors, who suggest greater "frame control" with enabling tools such as just-in-time to respect people on the frontline and respect the facts they share about what is happening to them. "Mastering the path as opposed to being led by it, means looking up frequently to reevaluate both destination and way as new information comes to light."
5S, Hygiene, and Healthy Habits
5S-like practice can uncover hidden beliefs and misconceptions, and pave the way to adopting new hygiene practices – as opposed to arbitrary imposition, argues Michael Balle, adding: In this community, we, of all people, have been trained to do so. Now is the time to start acting on it.