Organizations of all sizes are involved in business planning. This includes setting ambitious targets on sales and other KPIs to drive growth and efficiencies. This post covers how using data on future events can help your businesses improve month-on-month planning accuracy through leveraging Demand Sensing Analytics & Machine Learning.
The retail industry is a leading example of how events can impact sales and supply chain dynamics. For example, sporting goods companies need to plan their sales around World Cup and the Olympics. Such events have a drastic impact on customer behavior, and companies need to find a way to take this into account when taking supply chain decisions.
The diversity of events is also large. There are numerous events both outside and inside the organization that affect your P&L:
The diversity of events creates challenges for the planners to come up with accurate financial forecasts that take events into account. It is difficult to estimate how much impact a certain event will on future sales. For that reason, automated systems leveraging Machine Learning emerge to help planners take the relevant events into account.
Specifically, 2 types of events are important to business planning:
In both cases, having a system that is able to automatically take into account the relevant future events allows planners to increase the accuracy of the future sales targets.
Recently, Agnicio completed a project for a large retail client that involved identifying events’ impact on sales planning. The goal was to find out how can we successfully leverage events data for our overall Demand Sensing solution.
This exploration resulted in the development of a new “events driver” which became the latest addition to Agnicio’s framework of demand sensing algorithms. In the remainder of the blog post, we cover how we approached this task.
Our data science team followed a 3-step approach to improve the forecasting accuracy of our Demand Sensing Solution that can forecast sales 18-24 months ahead on the country, store location, product category, and product line levels.
Steps involved:
The first step was to come up with a list of public and company-specific events that could be relevant to the organization. We overlayed event dates with daily sales data to match against the observed peaks or drops in sales.
This allowed us to have a conversation with the business planners and identify which of the fluctuations in sales are due to the events. The key metric we optimized for is the percentage of the peaks/drops in sales that we can match to our events list on a given year. The outcome of the workshop was a list of events per country that can be leveraged by the data science models.
The next step after that was to identify for how long a certain event has an impact on sales. When doing monthly planning, it is not enough to say that an event happens on a certain month. If customers start shopping 3 weeks before a giveaway campaign that falls on March 1st, should the sales be allocated to March or April? In that case, using event weights allowed us to split the event impact between multiple months around the event.
To find the event windows we used a Neural Prophet model released in 2020 by the Facebook research team. While we do overall planning on a monthly level, daily data allows us to see on a day-by-day basis when a certain event starts and stops having an impact on sales.
Here is a simple example from Facebook how Super Bowl has a stronger effect on the day before compared to the day after the event:
With this approach, it becomes clear when an event impact starts and ends.
To arrive at the final result we integrated events in the Agnicio Demand Sensing framework. Then we used multiple event combinations to test accuracy at different forecast periods up to 18 months forward.
In reality, we saw that reducing the number of events to only key relevant events gave us the best accuracy improvements. The model showed up to a 5% improvement in accuracy, which is very good given the overall model accuracy ranges between 80-95%.
Adding events to our Demand Sensing solution also allowed us to isolate the negative impact of certain extreme events such as Covid-19. As you know, Machine Learning models learn how to predict the future by looking at the past. In the case of the 2021-2022 forecast, using 2020 as a learning period makes the model biased due to covid lockdown measures.
To correct this issue we added a ‘covid’ event into the model, which allowed the model to learn that what happened in the past is abnormal.
In case you’re wondering if the planning team can add their own events to the events driver, the short answer is yes. We’ve been able to successfully demo that approach in the new SupplyFocus tool in collaboration with our partner Cubewise.
This approach has several advantages. First, planners can mark certain dates in the past as events, so the driver learns their impact. Then it only needs future event dates to extrapolate the effect. In case an event has never happened before, a manual indication of expected impact can be provided. After the event has passed, the driver will look at actuals and learn its true effect. Then the planner can decide to apply the event again or mark it as a one-off event.
Using events interactively requires some data management on the client side, but brings additional benefits in terms of improved planning accuracy and agility.
Is your business impacted by events? Do you have a large planners team involved in business planning which creates a large overhead and costs? Would you like to use the latest AI technology to improve business planning?
Book a Demand Sensing demo today by reaching out to info@agnicio.com or by connecting on LinkedIn to the CEO of Agnicio: https://www.linkedin.com/in/hichamelarfaoui/
Let machine learning help your company re-imagine business planning to drive growth and achieve operational excellence
We help companies to drive sustainable growth and achieve operational excellence through data analytics strategies & solutions