If Amazon materialized today out of thin air, few sane people would feel confident about being able to run it profitably.
Globally, the company sells more than 3 billion products through 11 different country marketplaces. In the U.S. alone, Amazon introduced 208 million new products in 2018 — most of which are slow movers, or ‘long-tail’ items. Yet despite its extreme complexity and scale, Amazon’s July, 2018 earnings doubled shareholder expectations, turning a quarterly profit of a whopping $2.5bn.
Amazon’s success might seem like a fluke that’s impossible to replicate. However, when you break it down, Jeff Bezos became the world’s richest human by figuring out how to do three seemingly contradictory things at once: deliver exceptional service levels, at the lowest cost possible, and manage complexity. I argue that achieving all three with true success is difficult for any company to obtain without advanced technology capabilities.
First, change your mindset. Unfortunately, in a market of growing demand volatility and higher service expectations, too many companies get locked into traditional processes and a familiar downward spiral. Unable to reliably forecast an increasing number of SKU combinations, they load up on inventory to accommodate long-tail, erratic demand. This invariably leads to problems like extra freight costs and excess and obsolete inventory that either needs to be written off or sold at a heavy discount. Planners are continuously in reactive and inefficient “firefighting” mode, spending most of their time changing suggested replenishments and manipulating service levels instead of driving performance.
Supply chain problems are often hardest to fix because they have counterintuitive solutions. If you want to achieve “Amazonian” success, the first thing you must do is get out of your own way and try a new approach. As the well-worn adage goes, the sure sign of insanity is doing the same thing over and over again and expecting different results.
The secret to service-driven planning is probability forecasting and stock mix optimization. The way to manage complexity and achieve high service levels is to first break through the forecast accuracy barrier: Instead of forecasting one number, understand the range of possibilities of demand in your forecast. This method is called probability forecasting. Using this approach, you still get one number that’s associated with the most probable outcome. However, banded around this number you get a range of other possible outcomes, each with a different probability attached. It’s an alternative to traditional “one-number” forecasting, which is based on averaging aggregate order history numbers.
Probability forecasting is ideal for supply chains that include a high number of long-tail items and face demand variability and uncertainty due to the huge number of factors they are unable to model adequately. As supply chain guru Lora Cecere concluded in her blog Probabilistic Forecasting: Right Fit for Your Business?: “For difficult demand profiles, probabilistic forecasting is a new and powerful technique. It is a type of engine. Forecasting is all about better math, and the fit of the data model to drive outcomes.”
Your supply chain doesn’t have to get too complex before it can benefit from probability forecasting. Here’s an oversimplified example to illustrate. Say you want to forecast demand for a specific car tire’s SKU. A single-number forecasting system would look at this tire’s history of selling four units per month and identify average demand as one tire per week. Because this forecast does not address customers replacing all four tires at once, it would continually propose the wrong forecast and hence inventory levels to meet target service levels.
For inventory planning, you need to know the probability of each line-order quantity — for one tire, two tires, three tires, four tires, etc. Probabilistic forecasting provides exactly that information, identifying the order patterns (e.g., order size, order frequency) that inventory can use to service demand.
We cut a lonely figure in the supply chain world back in the early 1990s, when we first began to champion this approach as an alternative to traditional forecasting. Today, faced with overwhelming supply-chain complexity, there’s more of a burning platform for companies to try it out. Invariably the feedback we get is “I only wish we had started doing this earlier!”
Secondarily, stock mix optimization enables what we call “service-driven planning,” by taking advantage of the scale and variability — the complexity — of your SKU portfolio across the network. Instead of assigning the same service level for every SKU in a group, each SKU location across the supply chain is assigned its own service level that is optimized to achieve the business goals. For instance, instead of assigning all SKUs in a class a 98-percent service level, a global 98-percent target is achieved by optimally setting individual SKU location service levels at 99 percent, 97 percent, 99.5 percent, etc., achieving the same overall customer service level objective with far less inventory expense.
Global prescription lens manufacturer Shamir Optical applied probability-based forecasting to become more service-driven. Rather than use a one-size-fits-all inventory policy, Shamir analyzed demand patterns to create a blend of different service level targets for each individual SKU in each location. The firm reduced inventory levels by more than 25 percent overall, while consistently achieving service levels exceeding 99 percent.
Probabilistic forecasting can’t practically be executed fast enough by human planners. To make it work, you need to automate the planning process with a self-adaptive system that uses machine learning technology — a form of artificial intelligence. To generate probability forecasts, you first need to model your supply chain. Most companies start with a sample group of SKUs to test and scale up over time. Against your model, you then need to factor in the impact of a wide range of potential demand variables. These can be traditional inputs such as order history, other corporate sources like customer relationship management (CRM) system data, and even external sources such as weather, stock market and social media trends. Because a supply chain model is a “living” system, machine learning continuously learns and tunes the results over time, allowing you to introduce new data sources as needed. Applying A.I. provides deep insight into the behavior of demand and inventory to improve the outcomes.
The good news for humans, however, is that probability forecasts are, by design, a starting point — not an end game. They are designed to give planners the data they need in time to make informed judgement calls on service policies and corresponding optimal inventory levels across their supply chains.
This A.I.-augmented probability forecasting represents the ideal symbiosis between human and machines. The system gets smarter over time by factoring in human input, and the humans get smarter by learning from the success rate of the probability forecasts. This frees up planners to focus on service, work on strategic projects and add their business insights to the system.
Your business can thrive on complexity, too. The beauty of probability forecasting is that while service levels go up, costs, waste and inefficiency go down. Hundreds of companies like Shamir Optical have gained a wide range of benefits, from freeing up working capital to reducing obsolescence, transportation and expediting costs and markdowns. Many companies report becoming more responsive to market changes and being able to make better strategic decisions.
For people wed to the “one-number” deterministic approach, probability forecasting will feel counterintuitive. However, unless you are in a commodity business with few items and totally predictable demand, the one-number approach won’t cut it. Amazon not only uses this method, it also offers a probability forecasting tool to partner vendors. Isn’t it time you gave it a try?
Joseph Shamir is chief executive officer of ToolsGroup.