By Sathish Kumar, Solution Architect, Industry Cloud & Retail, SAP Labs India
In today’s world, retailers and manufacturers must go through tremendous competition, fast changing trends, social media influence etc. to stay in business and be profitable. Many companies are struggling to find the right combination to survive. The key to run a financially healthy company today is Artificial Intelligence / Machine Learning (AI-ML) in certain decision-making processes.
AI-ML concepts are gaining momentum and can be applied to various components of the Supply Chain – Demand Forecasting, Logistics & Transportation, Inventory Management, Promotions etc. The Key is to have the right product at the right store at the right time and in right quantities. In this blog, we will focus on usage of AI-ML concepts in Demand Forecasting and how this helps retailers and manufacturers better plan their assortments in store/online.
Why Demand Forecasting? It is fundamentally the process of developing an estimate of the future customer demand. It uses Customer Data like historical Inventory, Sales, Current Trends, Calendar events like festivals and Sports seasons to estimate and predict Customers future demand for a product. It helps reduce risks and make efficient financial decisions like profit margins, expansion opportunities to effect increased sales in the future and overall improve the health of the supply chain. Not just retailers but this also helps manufacturers on what to focus on the near future.
So, how is this done? There are few successful products in the market which offers demand modeling and forecasting services for applications driven by demand prediction. It also provides insight into shopper behavior, enabling retailers to perform predictive analytics on consumer demand. There are several forecasting methods used like ARIMA Models, Exponential Smoothing, Time Series regression models. Here, let’s look at some of the core pillars and data that is commonly used in all the models.
• Demand Influencing Factors (DIF) – These are the factors that determine the demand for a product and importance of each factor differs by the product. Some of the factors include Income Distribution, Sports / Holiday season, Taste preference, Population around the store, Prices of the Product, Competitor Prices, local events etc. The importance of these factors can be determined by the forecasting algorithm in the modelling process using historical trends.
• Hierarchical Priors – Future trend of the product also depends on the past sales history of similar products. As we know, trends keep changing because of various reasons and this needs to be factored in. Consider an average of 5 quantities of XYZ were sold last year does not mean that it will sell the same this year as well. This is where “Priors” come into picture. “Priors” help us to determine how much of this historical trend needs to be factored in forecasting future demand. So, what is hierarchical priors? Every product belongs to a category / sub-category of products naturally forming a hierarchy. This hierarchy has an influence on the product selection. So, we naturally have a hierarchy in the priors. These priors can come from historical sales data, corresponding product categories, similar products, past trends etc.
• Product Attributes – Product Attributes play a major role in the sale of products. During Christmas season, the color “red” is highly sought after and during Diwali season “Decorative” items are highly sought after – to just say an example. These attributes are generally weighted, and they are also given as an input to the Forecasting algorithms. These weights could change time to time. What is an Attribute? – Any characteristic of a product that can be used to explain the product can be considered an attribute. It could be categorical or ordinal or could be intervals. Companies can look for “best sold” attributes in the historical sales and aim to use similar products in the future for the same period.
• Demand Decomposition – Explaining “why” an ML model determined a certain result is the toughest task to do. But this is the most important aspect for a customer or an end user to understand and follow the system recommendations. It indicates how much of the total value can be attributed to baseline business and how much is due to specific DIFs. A DIF can have a positive or a negative influence on the demand. Examples include offers, prices, public holidays, seasonality etc.
Until now, we talked about the variables needed to execute forecasting and the model parameters. But the output of the model is also equally important to analyze. In a general Machine Learning solution, visualization is key to understanding. A well written ML algorithm may lose its importance if it is not properly explained to the customer visually.
Visualization is generally provided where customers / end users can slice and dice the data to understand how forecast values compare with the historical values and can see how the DIFs influence the values. There could be many such visualizations required to understand the results.
The Visualization generally provides a good idea for the end user. And based on the user’s experience and decision by management he/she can accept the results or tweak a little to fit the overall needs.
The world is highly interconnected right now, and trends keep changing faster than the companies could adapt. We rely heavily on social media, social media trends also become one of the influencing factors for forecasting. Every company / user uses such trends in a proprietary way to enrich the result. Hope the existing forecasting solutions also adopt to the changing world, so the forecasts can be more accurate, so as an end consumer we get the right products at the right price.