To determine future demand for the products you sell, you can either employ judgemental forecasting techniques (e.g. “I think we’ll sell X units”) or you can use quantitative forecasting techniques, which use historical sales data to generate demand forecasts using a variety of statistical techniques.
Quantitative methods certainly offer many advantages. The resulting demand forecasts are free from human bias and statistical algorithms are often able to identify patterns that might not be obvious otherwise (for example, a good forecasting package will be able to discern between a logarithmic and a linear trend or between additive and and multiplicative seasonality). As such you should try to employ quantitative forecasting methods wherever possible, even if only to establish initial demand forecasts (that you tweak later using your business expertise).
However, generating statistical demand forecasts can be very computationally expensive. Even on a powerful computer, it might take anywhere from 3-10 seconds to fit parameters and run a variety of algorithms, pick the best algorithm and finally generate a demand forecast for an individual SKU. This doesn’t sound too bad, however that’s just a single SKU and even small retailers regularly carry tens of thousands of SKUs. So building demand forecasts for all of a company’s products on a desktop computer might take hours, days or even weeks!
Scotty: I’m giving her all she’s got, Captain!
The most obvious solution is simply to throw muscle at the problem, scale up and buy a powerful server to run your forecasting software on. However, powerful servers don’t come cheap and generally a faster CPU and a bit of RAM is only going to make your forecasts marginally faster… maybe by a factor of 10, if you splash out on something resembling a super computer.
For companies that manage lots of different products or product variations then, this solution is expensive and not particularly effective.
James T. Kirk: All she’s got isn’t good enough! What else ya got?
A better alternative is to scale out horizontally and have multiple cheap forecasting servers generate demand forecasts for your products concurrently.
A good way to explain why this works is to think about Lamborghinis vs. Toyotas1:
|2011 Lamborghini Murcielago||2011 Toyota Yaris|
For around the same price as a 650 horse power Murcielago, you could buy 30 Yarises with a combined total of 3900 horse power… so roughly six times more cost effective than the “vertical scaling” solution.
The figures obviously change a bit when you start comparing a Super Computer with a Raspberry Pi, but the same basic principles hold and, overall, scaling out horizontally turns out to be much more cost effective than scaling up vertically.
SkuBrain Forecasting takes horizontal scaling one step further and scales forecasting jobs out into the cloud.
With the advent of cloud computing frameworks such as Amazon Elastic Compute Cloud (EC2) and Windows Azure, it is now possible to lease armies of low cost computers by the hour to work on big forecasting jobs without needing to buy or manage any hardware.
Need to generate demand forecasts for 100K SKUs? No problem – we’ll just lease 100 compute instances from Amazon! When your forecasting job is done we shut the instances down and somebody else can start using them (and more importantly, somebody else can start paying for them).
By renting the computing power that we use in the cloud only when we need it, we’re able to provide an extremely powerful demand forecasting and replenishment planning solution for a fraction of the price that you pay for traditional supply chain management platforms. And, as with all good cloud based applications, you don’t need to buy or manage servers, there are no “implementation” or “consulting” fees and we’re not ashamed to publish the pricing on our website.
If your current demand planning solution is a bit sluggish then, it might be time to move to the cloud with SkuBrain!