User Centric Product Design in a Machine-First Demand Planning Software
Demand planning starts with a machine forecast but it does not complete without the addition of human insights, layering on inventory strategies, supporting consensus planning amongst others.
A highly automated, reliable, accurate machine generated forecast is a great starting point of a demand planning software. In a mature state, this machine forecasts get more and more accurate for a larger dataset, resulting in saving demand planner time for strategic and collaboration work. The demand planner role changes to someone who educates the machine forecast to continuously get better.
To put forecast management complexity in perspective, in a typical retail retail company, there are 30–50 demand planners managing 90–150 million SKUs. Per rough math, it translates to 3 million SKUs per planner. While the goal is to get as many of the machine forecasts perfect, there are valid scenarios and/or forecast outliers and will need human intervention.
With traditional demand planning software, the number forecast errors that require a demand planner attention are tens of thousands per week, the number of inputs that the models can handle are very limited having the planner to hand key and/or change a lot of forecast numbers. Roughly 60% of the forecasts are touched by a demand planner.
In a machine-first demand planning software, two things need to happen
(a)The machine forecast needs to get much better by leveraging AI/ML and Neural network like algorithms additionally. This topic is covered in detail in a separate article. https://medium.com/@savithanallasamy/checklist-for-a-differentiated-demand-forecast-engine-75ed367051c3
(b) The user interface(UI) for demand planner needs to be extremely intuitive to pull in planner/user attention to things that machine needs support with as well as make forecast management easier in order to drive improved user productivity.
A mature demand planning software produces the best possible machine forecast, augments it with human intelligence to predict when and where the customer demand is. It is focused not just on forecast accuracy metrics, rather driving improved inventory position measured by in-stock improvement, safety stock reduction, lost sales reduction, excess inventory reduction.
Here are four key categories of UI capabilities needed in a demand planning software
#1 Allow human to easily review and augment the machine forecast, “educate” the model on what it is potentially missed out.
#2 Provide automated exception detection to catch, correct forecasts before they are actioned. It is important to catch a bad forecast from sending too much strawberries to a store or too less bananas to a store or ordering too much and then realizing the demand is just half of what has been ordered.
#3 Align the demand forecast to fulfillment and sales strategy
#4 Enable full transparency to forecast metics and the impacted inventory metrics, as well as callouts to potential inventory $ saving opportunities. After all, what is not measured and tracked cannot be improved
A best-in-class demand planning sofware accomplishes these objectives through capabilities, a few key ones listed below:
- Predetermines, presents a pre-prioritized list of outlier forecasts that need human review and intervention. Call for action, list of exceptions with recommendations. Examples could be forecasts that are 10x any historical sales. Another example is a new item’s forecast created with no or very limited history and with very low confidence.Another one could be a location whose forecast has been consistently biased over 20% over last 8 weeks.
- Clear workflow for outlier forecasts along with who is accountable, how soon it is expected to be fixed, if the outliers have been repeating etc. Additionally, recommending alternate solutions might be the next step in maturing Ex. Switch to a simple weighted average calculation for this item1-location 1 forecast until there is 12 continuous weeks of non-zero sales.
- Makes forecast review and management extremely organized and intuitive. The most proven approach is to to support product classifications/groupings. Product classifications could be pre-created with the additional option of allowing the user to to create custom cohorts of products. Additionally, what is neeeded is a easy to pick, sort, filter, and tagging mechanisms to choose product(s) and/or locations based on velocity of sale, units/sales contribution, seasonality, event tagging.
- Provides multiple levels of forecast controls for planners. This is needed to manage the volatility of the forecast, especially when a demand planner manages thousands of store SKUs.
- Built to support and automate the consensus planning process. In a traditional setup, Demand planning teams spend 20–30% of time doing consensus planning and manually adjusting the forecasts for consensus discussions. An example of internal consensus forecasts is (a) Finance team creates a $ forecast, which drives communication to Wall Street. This forecast is also an indication of the company’s investments in select brands and types of products. (b) Merchandising teams set sales targets and plan buying to to meet those targets especially for new products/product lines. (c) External consensus forecast with vendors gives retailer access to vendor’s understanding of color, style and market as well as insights to vendor calendar for new lanunces, vendor funded promotion calendar. The number of interactions and manual forecast updates calls for efficiency. What is needed in the next gen planning software is for “alternate” forecasts from various internal and external planning and collaboration processes to flow into the planning software, a sophisticaed rule engine that prioritizes “alternate” /consensus forecast over the machine forecast based on type/stage of the product based on business rules.
- Provides ability to add context to both internal and external data inputs before being fed to forecast engine. Ex. Placement of a product in a store has a huge impact on demand for many products, especially for impulsive buying products. Having a systematic planogram input into the model is not enough, we need to know is say, Aisle 21 is a hotspot or not. Sometimes these are qualitative inputs that need to be augmented to the data.
- Provides planner the option to review direct and indirect external causal factors for a forecast. Ex. Next week weather, hurricane forecast and impacted areas, commodity prices, any social media trends related to the product/category/related categories, upcoming holidays, local events, upcoming competitor locations.
- Provides ability for demand planner to “block” select products from the effect of global forecast model effects. Ex. Post pandemic period, some products continued to the same demand trend, some slightly close and come back to the pre-pandemic period. Select product categories might need to be excluded from the model’s “pandemic correction”
- Enables a two-way communication between human and forecast engine. Allows the user to “educate” the model that the strawberry picking problems in a particular region is what caused the low sales that year. The model in turn learns to ignore that region’s sales for that year and instead look at the other years’ sales trend.
- Layering the stocking and/or marketing strategy on the machine forecast: An example is the ability to adjust the forecast to support different location and/or market strategies. Ex. If SouthEast states are a growth market, the company might invest in market specific strategies (ads, marketing events, partnerships) that would drive demand and would also need. Another example — Before the start of a high sales period (promotional week, labor day) for a select set of product cohorts, keep the inventory high to support sales strategy
- Provides ability to overlay the supply chain fulfillment strategy. Ex. ability for the supply chain team to map the historical node to “desired” stocking node. A customer demand might exist at zip code 12345 and zip code 34567. Mapping the demand to the “optimal” desired fulfillment node is a combination of science, art and strategy.
- Calls out opportunities to reduce safety stock. Ex. If an item-location forecast has consistently had a good accuracy and a tighter bias, then the safety stock setting could potentially be revisited and possibly be reduced
- Built to support the business review processes. In many companies, a weekly inventory review is a common cadence. Most often when the order projections are higher or lower than “expected” and when attributed to a demand forecast , the planner is asked to explain what makes up the forecast. Being able to explain “what makes up the forecast number” (explainability, breakdown of forecast units) is a critical expectation of a planner
- Provides full visibility to the quality of the forecast over a period of time. Ex visibility to last 52 weeks of forecast vs actuals, Machine forecast accuracy vs planner adjusted forecast accuracy and seperately Planner value add (also referred to as Forecast Value add). It also tells how good the science-based forecast is, how good the human input and how good is the final executed forecast. It is also important to define and track meaningful forecast metrics . Ex. For sets of items that sell less than 1 a week, it makes to measure the forecast at a month or quarter rather than weekly.
- Provides visibility to inventory metrics correlated with forecast. What was forecasted to be ordered vs what was order quantity, what was forecasted to be replenished vs what was replenished quantity, differences between what was forecasted, what was ordered and what was replenished and if they are attributed to the variability at the time of ordering vs time of replenishment. Additional visibility to lost sales, excess inventory, instock %, Out of Stock Days/Weeks will aid decision making.
- Publishes the supporting forecast data needed for rest of inventory planning decisions like safety stock, stocking strategies i.e, which locations to stock, which location to fulfill in a constrained supply situation. Another example is to be able to provide a forecast segregated by type of demand — Pick up In strike demand, Direct to Customer demand
- Becomes the single source of demand forecast for the enterprise not just for inventory planning but for distribution center labor planning, transportation planning, vendors. These additional use cases need that the demand forecast extend to additional Unit of Measures like Cube, Weight, Number of trucks, Labor hours, etc
Summary
The perfect blend of continuously improving forecast generated from the best of forecasting algorithms/techniques combined with the human insights is what makes a demand plan reliable and subsequently drives the best inventory position and ultimately drives customer satisfaction and company profitability.
Related Stories