Branch Ranging II, a case study

In addition to its wholesale trade the company - one of the most diligent inventory science practitioners in the world - ran a network of well stocked branches.

The company range was over 250,000 SKUs (product lines).
A branch might stock one tenth of the full range.
Beyond a few fast movers, branch lines move at a glacial rate.
Even the fastest movers only sell once a branch per day.

The company were applying some techniques they had honed on fast and medium movers, but doing so at branch level.
They were constantly operating below the Threshold of Forecastability (TOF), the rate of sale below which any reforecast will more likely do harm than good.

In the Branch Operations Case Study we looked at the knock-on problems reforecasting was causing. Here we look at TOF …

Setting stocks without forecasting

Well, that's a challenging title!

Luckily, there are some breaks …

1. A well tuned supply chain is hugely resilient to forecast error. If the same stock minimum is appropriate for all sales between 7 and 27, then we gain nothing by reforecasting sales in the middle of that range.
2. Most supply chain are kept in stock by batch (pick) quantity far more than they are by science.
   Indeed, we have unpicked supply chains with terrible science yet adequate performance; the 'get out of jail free' card being batch quantity.
   Turning batch quantity into a virtue provides a cheat escape from reforecasting the un-forecastable, aka 'maths gone mad'.
3. All supply chains trade off effort vs. 'good enough'.
   The corollary is that being a bit wrong some of the time is acceptable. It beats the hell out of being a lot wrong all the time.
   I speculate that we want forecasting to be our magic bullet. Worse, the more enamoured we are of the science, the more we want to believe it's the only bullet. [1] Which it never is and - below TOF - never can be.
   So 'less wrong more often' is as good as it gets.

'Not forecasting' is a tabloid headline, alluring but stupid. To set some stocks at 2 because we believe they will sell more than those with stocks of 1 implies a forecast, however derived.
Something made us believe the 2 would sell out more often than the one … and that's a forecast!

At issue is how we make that call when both products are below TOF.
Further, even though both cannot be forecast without risk of misleading us (one colloquial way of reading TOF) we'd really like to know if a stock of 1 was due promotion to a 2, or if a 2 was due demotion to a 1
We'd like to know if the forecast changed which, in a complete Catch 22, means we want to reforecast those which have changed. And the only way to do that is to reforecast everything then sift those thought to have changed from those thought not to have changed, which rather brings us full circle!

At heart is that we don't have enough fresh evidence to re-forecast.
Might there be fresh evidence elsewhere. Before we answer that, let's take an excursion.

There's a famous correlation, part of statistical folklore, between the number of convictions for drunkenness and the number of methodist ministers.

We are meant to fall into a trap, that more teetotal preachers cause more drunks. Or more drunks become ministers? Of course the common cause was rising wealth, which generated more ministers and more drunks.

But suppose I were a betting man who knew about the rising number of acolytes and was offered an evens bet on whether drink convictions would rise or fall.

I'd bet on a rise; I'd be stupid not to!
The point is that, for a correlation to be useful it doesn't necessarily have to be cause and effect, it only has to be consistent.

Grey sky = rain, take an umbrella isn't actually a particularly good correlation (measure how often the brolly stays furled. And how often sunny days produce summer storms). But we use it, and are smart to do so.[2]

With modern software we can search to see if anything correlates with anything. It's important not to go overboard … a historic correlation with the number of lamp-posts would provoke derision. [3]
But there is a correlation, a very good one (88%), believable (i.e. logically defensible) and not quite where we might expect.

Details are not just client confidential, they are found in different places in different types of supply chain.

Suffice to say it will prompt a >10% increase in service level on less stock, with less effort once it is absorbed.

Why the last phrase?
This supply chain has been tuned by dedicated humans, pouring endless effort into the misleading methods they have been given. It's a fiction to imagine they will fully adopt a new method overnight. They need to be convinced emotionally and not just logically.
So the program anticipates giving them old and new choices in parallel. They will find over time that the new is better.

We might have a forecasting problem, please contact us

1. As a long time member of a forecasting SIG, speaker at the US Institute of Business Forecasting, inventor of a new method for forecasting slow movers, and manager of one of the first focus forecasting systems, I've been there.
   When the T shirt fits I'll happily recommend it.
   It doesn't. Don't touch it with a bargepole.
2. Like shop inventory, it's a bit more complex than 'got wet' vs 'stayed dry'.
   We're balancing the higher inconvenience (a 'cost') of getting soaked against the lesser 'penalty' of taking the brolly unnecessarily.
   The point stands, we're doing so based on a pretty scruffy correlation.
3. No surprise that the number of lampposts also reflected rising wealth, and so correlated well with drunkenness.