We offer the core supply chain insights and technologies, including:-
SLIT,
the definitive tool for balancing ServiceLevel, Inventory & leadTime.
Fix any 2 variables and the third is then a consequence.
Visual Control for the Supply Chain (VCSC) animates any supply chain under a wide variety of settings as an aid to diagnosis, design and training
FireResponse shows how to meet service targets under erratic demand. (also works for taxis, etc) The related FlowBalancer works with varying and unpredictable cycle times
Why we founded Supply Chain Tools
Supply Chain Tools is dedicated to bringing the core tools and techniques to wider use. As each new supply chain insight is developed, we take time to encapsulate that learning into a recyclable solution.
Does this mean you have a ready made solution to my problem?
No idea! We don't know what your problem is yet. While
you may be able to describe the symptom, only finding and fixing the underlying
problem will give lasting benefit.
But once we do find and agree the cause, we don't prescribe surgery if aspirin
will do.
Recycling means that where a tool like VCSC forms part of the solution,
we do it very much quicker and cheaper. The core tool will still need some
customisation - 'part packaged' describes the approach, rather like buying
turf for the lawn.
SLIT is the definitive tool for trading off Service Level, Inventory and lead Time.
Fix any 2 inputs and the third is a consequence. For example
And if we don't like the consequence, we must change an input typically T
The reference site has achieved nearly 50% inventory reduction
(on $1.9m), while emergency shipments (as their rough and ready measure
of service level) went down by 2/3rds.
With more to come.
Other uses have included
Above all, SLIT provides the tools for defending necessary stock, and removing unnecessary stock. Because the trade off is a science, we can now also gain consensus between sales, accountants and operations.
SLIT supports both fair forecasts and those with 'unwitting bias'. Any forecast is derived from a pool of potential 'buys' of which actual demand is merely a sample. With slow movers (everything moves slowly towards the bottom of the supply chain) the error in deriving the forecast is greater than the error in believing it.
There's a brochure and full demonstration available. Just call. To Top >>
VCSC simulates and animates any supply chain as an aid to diagnosis, design or learning. It drags real demand data through a fully featured electronic reconstruction of the supply chain. Your demand, through your chain, at your chosen settings, if you will. But with a risk free answer in 27 seconds, not 13 weeks (an early run was on NZ wine)
Settings include forecast method and smoothing (also forecastless)
seasonality method and smoothing (including on/off), ROP (or similar) method
and setting.
Within ROP or BTL are the SS and EOQ and variants. Then there's re-order
review frequency, lead times, and the penalty costs of overstock, understock
and out of stock.
That there are ~20,000 possible combinations of input settings
puts the logistics task in perspective. Optimisation is a red herring, a
complete pipe dream.
Success is about understanding cause and effect, and managing the big causes.
VCSC helps users, managers and directors understand the supply chain, the
dangers of hindsight, and the need to (and cost of) reconcile conflicting
objectives.
The product is it's own best advert, just call for a demostration.
VCSC takes the customer's own demand data (or any other data) and shows how a supply chain would react under a range of about 20,000 possible input settings
The inputs include
Outcomes include anything which can be known from such a model. The list is therefore limited only by our imagination and need, but has typically included
VAlpha (Variable Alpha) gives better forecasts than previous methods. It's simple and intuitive.
Most of the difficulties with automatic forecasting occur on slow movers. Most products move slowly at some point in the supply chain.
The benefits of using VAlpha are:-
A much better forecast for slow movers.
A better forecast for fast movers.
An acceptable forecast for medium movers.
With adaption, a better forecast for new products.
For many systems, a very simple program change.
How is VAlpha different?
Most systems re-forecast based on the latest month's demand data. The business and planning / purchasing cycles are then built around the 'new' forecast. Yet something selling 100 times a month has 100 times more meaningful new data as a product selling 1 a month. It may be that the fast mover should be re-forecast weekly, and the slow mover only yearly.
In other words, re-forecast only when there is meaningful new data.
Intuition confirms this. If a product line sells one a month, demands of zero, 2 or even 3 seem 'quite likely'. Yet if it were selling a regular 100 a month, it would be surprising if it sold less than 85 or more than 115. For slow movers, a forecast error of 100% seems 'quite likely', yet for fast movers an error of 15% seems 'quite unlikely'.
Yet most forecasting systems either treat slow and fast movers the same, or are so complex that only the originator can maintain them.
Forecasting systems tuned for a mid range product are too responsive for slow movers, and not responsive enough for fast movers. VAlpha adapts Poisson's 'chance event' theory to predict the best weighting factor to use for each and every product line.
New Products
The average retail product life cycle is now less than 3 years. Waiting for history risks missing the boat.
A system which works with limited history will deliver
clear benefits. Standard VAlpha can be modified to fast initialise - in
sectors like electronics and fashion even to accentuate (upweight) the early
demand data.
See paper for more detail >>
The "Threshold of Forecastability" is that point
at which the error is always ±100% or greater. This figure, which
varies from case to case, is much higher than previously thought.
A range of solutions and workarounds are available, depending on circumstances.
For small numbers, "the error in deriving the forecast is greater than
the error in believing it"
Common Symptoms
If stock is often in the wrong place, if you find you are buying stock you
already have, if DRP (Distribution Requirements Planning) appears unstable
or illogical, it may be that you are trying to forecast numbers which are
'unforecastable'.
Beware, a very large corporation nearly went bankrupt through this problem.
Another, smaller orgaisation recognised the warning signs from our discussion,
and averted the problem. To Top >>
Here's an extended summary of the paper I presented at the Institute of Business Forecasting conference in Chicago. The subject was VAlpha (Variable Alpha - a new method of demand forecasting) which gives much better forecasts on slow movers and better forecasts on fast movers. It's prime goal is better supply chain performance. Forecasts are just one input to the supply chain (and often only the 4th or 5th most important) but better forecasts are still better inputs.
Coincidentally Warwick's Dr. Richard Wilding had started a hare running (in IoL's Focus and elsewhere) that software could cause chaos in the supply chain. He identified a problem which VAlpha solves, or partially solves. In addition, VAlpha solves some other problems inherent in small number behaviour which occur in almost all supply chains.
This note, based on mine to Richard W, brings the two strands together.
You are free to circulate &/or use VAlpha.
Anyone is, I put it in the public domain deliberately to speed uptake.
If you do use it, I expect you to acknowledge its origin.
VAlpha was developed specifically to damp erratic supply chain behaviour, which may (or may not) be the same as chaotic. My objective was an empirical solution, deriving an equation from something without (necessarily) needing to know why it worked.
Which gave my track chairman the chance to tell of the economist who, faced with a perfect and proven empirical solution, said "Ah, yes, but does it work in theory?"
Each section starts with a recap to make sure we are talking about the same things; my own narrative then follows.
RW talked about 3 types of supply chain distortion, Forrester, Chaotic and parallel (which I call collateral)
Taking the last first, collateral damage is caused when the factory, and thereby suppliers A though Y shut down through shortage of components supplied by Z. The simplest case is a strike - Z being the supply of labour. The worse the failing supply chain is relative to the others, the worse the collateral damage.
Apart from building the best supply chains - including eliminating Forrester and chaotic effects - there's nothing more the designer or manager can do.
Second, Forrester
Forrester particularly affects overstuffed multi level supply chains with a single customer and 'opaque' layers where there's no collaboration between customer and supplier. Lean supply chains, 'pull' supply chains, those with many outlets/customers, those with visibility or information sharing between layers, and imprest systems are or can be largely Forrester proof.
In other words, almost all (non-Russian) supply chains.
That Forrester is still talked about in multi customer/outlet supply chains (where swings offset roundabouts) is probably more a shortcoming of the modelling techniques than the supply chain design.
Colloquial Chaos Types - my interpretation
There's textbook chaos, out-of-control operation, and the appearance of chaos.
They may look the same to the user, yet solving the wrong problem can be fatal.
The appearance of chaos.
Stable, well designed supply chains can give unstable results when the minimum order quantity is large in comparison with the replenishment quantity. Experienced supply chain designers will spot this, and either live with the consequence, or alter the consequence by changing the case size, or by broken case picking.
Chaos introduced by out-of-control operation
Where the supply chain operation is out of control, either because the rules are not defined or not adhered to, or are defined but plain wrong, we get chaos. The last sometimes fits, and sometimes doesn't fit the textbook definition of chaos. The other 2 do not, they have (deterministic) causes, however illogical. All look chaotic to the users! All are surprisingly common. Here are examples of all 3:-
1. Manual store reordering was adding 100% to the variation seen by the DC, over and above that inherent in the sales as shown by the EPOS data.
2. Hoarding, often caused by shortage or rumours of shortage (sugar, toilet rolls) Accidental hoarding can be caused by large quantity discounts.
3. DRP (Distribution Requirements Planning) The solution is often as simple as switching forecasting off.
True chaos
I've made an assumption here, which is that forecastless supply chains are entirely deterministic. So if there is chaos, it must be in the feedback loop. Only two things feedback of which end stock is deterministic in any single lead time and the forecast is therefore the only (???) suspect. Chaos must originate in the feedback loop, and there is no feedback during a single lead time. i.e. no outcome affects an input to that cycle, only an input to the next cycle. After all my other work, I'd started with low number forecasting as the (only) problem. With that second assumption marked for revisit, let's see where the argument led.
If the forecast overreacts - it is too 'twitchy' - we can induce chaos into an otherwise stable supply chain. The symptoms are obvious, as is the solution - make the forecast less responsive, or (in many cases) simply turn it off. Here's a fun example of a forecast for, let's imagine, 9 different products
M-8 |
M-7 |
M-6 |
M-5 |
M-4 |
M-3 |
M-2 |
Last Month |
Forecast | |
| Item #1 | 0 |
0 |
0 |
1 |
3 |
0 |
2 |
4 |
2.4 |
| Item #2 | 0 |
1 |
1 |
1 |
2 |
1 |
2 |
0 |
0.9 |
| Item #3 | 0 |
1 |
1 |
0 |
1 |
2 |
1 |
2 |
1.5 |
| Item #4 | 1 |
2 |
2 |
1 |
1 |
0 |
0 |
1 |
0.7 |
| Item #5 | 0 |
2 |
1 |
0 |
0 |
1 |
0 |
1 |
0.6 |
| Item #6 | 1 |
0 |
3 |
2 |
0 |
1 |
0 |
0 |
0.4 |
| Item #7 | 2 |
1 |
0 |
1 |
1 |
0 |
0 |
1 |
0.6 |
| Item #8 | 1 |
2 |
2 |
2 |
0 |
0 |
2 |
1 |
1.1 |
| Item #9 | 2 |
1 |
0 |
2 |
1 |
1 |
0 |
0 |
0.4 |
So item #1 is likely to sell 6 times as many as item #9? Wrong! It sold 10 in 8 months; item #9 sold 7 in 8 months, which 30% difference hardly accounts for a 6 fold difference in the forecasts. On top of that, it's a trick. All the items have exactly the same underlying rate of sale, one a month. What has happened is that pure chance has sold 0, 0, 0, 1, 3, 0, 2, & 4 in the months leading up to the current forecast for product #1. The 40% smoothing factor in the forecast has then weighted the more recent months by 40% (Month minus 1), 24% (month minus 2) and so on. The preponderance of recent high sales is actually nothing more than chance - the chances of selling 0, 1, 2, 3 and more than 3 in any one month (at an underlying rate of sale of 1) are 37%, 37%, 18%, 6% and 2% respectively. However, the forecast has treated those demands as 'facts' rather than chance, and responded with a forecast 2.4 times the true rate of sale. Conversely, item #9 has a forecast 60% less than its true rate of sale.
Although this effect is intuitive (if I throw 6 dice I'll get an average of one six ... but I might well get 4 sixes in any one throw), and has been quantified for centuries, the existing forecast methods have coped inadequately or not at all. I surveyed most of the forecasting software packages in Chicago on the Friday of the conference - none cope properly. Yet in a modern system the forecast drives the supply and distribution chains whose outputs get actioned, without anyone really questioning how sound the basis for those actions.
It's for precisely this reason I invented VAlpha. This gives much better results on slow movers, and better results on fast movers. (Medium movers behave the same)
Some background, and a colloquial explanation of VAlpha
Most systems are set up to automatically reforecast every month. This makes sense where there is genuine new information (say when selling 100 a month - we've got up to 100 new pieces of information, which in total are unlikely to be more than 15% influenced by chance)
If we reforecast something which sells 1 a month, we only have one piece of new information, and the average error - ignoring sign - is over 100%
Where history contains such a high proportion of pure chance, this can induce supply chain chaos.
VAlpha damps the slow mover forecasts by giving each month's data a much lower weighting. The downweighting is decided by the prior forecast, NOT by the most recent month's data which might just be a 'lightning strike'.
In this, VAlpha differs from adaptive smoothing which works off the error.
VAlpha is predictive - it adopts the alpha which should fit.
Adaptive smoothing takes a shot at the alpha which, historically, would have fitted (this is a lay explanation, sufficient only for this narrative) and is susceptible, on small numbers, to misinterpreting chance as fact.
The best fit line suggests we should use higher alphas for fast movers. Although my original objective was to solve the small number problem, I now think capping alpha (commonly) at 0.3 has always been a compromise, and am convinced that a 'free range' alpha would give better performance across the whole supply chain.
Prevalence of slow movers
A very large supermarket chain volunteered that 90% of their product lines move less than 10 items per store a day. So even the fastest end of the supply chain has many real slow movers in it.
I showed mail order figures where in any 4 day period, 70% of the SKU's do not move. And large businesses where the median product only moves 8 a month. Glaciers outstrip some of the medium movers. Continental drift regularly finishes in the fastest 10 slow movers.
Summary
SLIT, which trades off ServiceLevel, Inventory and leadTime and gives the highest service level for the lowest stock, gives the leanest possible supply chain. A good place to start, and as Forrester - proof as it gets. (SLIT is part of a 'pull' supply chain) Where it's necessary to drive SLIT with a forecast (and the majority of occasions, forecastless is better) then VAlpha is not only the best tool anyway, it's as chaos-proof as it can be.
A bridge to management ... "fine theory, but what should I actually do?"
Or, more likely, a bridge to the systems designers? Most supply chains are run by 'black box' programs, some of which are pretty flaky and mostly the data is not well maintained.
We can only recognise chaos; we can't manage it directly, since it's a consequence.
Elimination requires that we recognise, understand, find and fix the cause(s).
If all the inputs are chaos proof it might still be possible to get an out of phase feedback loop (or 'hunting') through the forecast.
However, it's unnoticeable in all my modelling of multi-outlet supply chains and, even where it happens in the real world, there are easy, cheap, big impact fixes which should always be done first.
Like getting the controllable inputs into control, and using VAlpha and SLIT.
And if there is residual chaos, whether textbook or simply erratic, then concentrate on the pre-shortage alarms, and the reaction chain. To Top >>