The COVID-19 pandemic has fundamentally changed our daily life. The economic impact of the pandemic can be most widely visible in the consumer goods sector. From grocery and toilet paper hoarding to rapid demand increasing for Zoom and other video conferencing software, COVID-19 caused consumer behavior change and supply-chain disruptions. Manufacturers and retailers are trying to predict these changes in consumer behavior and forecast the future demand with various techniques.
Machine learning-based methods of demand forecasting in the retail industry leverage historical data, but the limitations of that data are clear as we grapple with COVID-19. The current state of customer demand has significantly changed, and the forecast accuracy will be diminished.
How can we adjust to predict the new demand paradigm? The following are three methods that may boost the accuracy of demand.
Short-Term POS Data Analysis
One proven and efficient technique for identifying shifts in patterns of demand is to analyze the most up-to-date point of sale data. The latest month or two months’ data can be combined with promotions, sales orders and upcoming shipments to make an earlier prediction of alterations in the behavior of customers.
Natural Language Processing Analysis
NLP (Natural Language Processing) techniques analyze news and social media to identify customers‘ mindsets through data mining and sentiment analysis. We can then zero in on what customers are buying most frequently, as well as their preferences and behavioral patterns through feedback in reaction to news events.
With a large enough sample size of customer feedback, we can use NLP modeling to notice changes in customer purchasing decision-making and identify goods quickly being bought out of stock.
Information Cascade Modeling
The technique of information cascade modeling lets us forecast in situations with stir and high demand. These ML modeling methods are capable of parsing the most current point of sale data articulating herd behavior to create forecasts with the cascade algorithm.
One example relevant to the current situation is how in the early days of the COVID-19 pandemic, consumers purchased huge quantities of toilet paper, as well as long-lasting and canned foodstuffs. Seeing others buying these items, still more consumers repeated these purchasing patterns. In response to significant increases in demand for these items, manufacturers ramped up production and, consequently, sales. This, in turn, produced an overabundance on the market, with supplies outstripping demand.
Data science engineers can use cascade modeling in e-commerce platforms to detect similar fluctuations in demand over short time frames, allowing for better and more accurate crisis demand projections.
Improving Demand Forecast Accuracy
Not every product can be accurately forecasted. This is the reason it’s so crucial to first determine how forecastable your product is before choosing one of the main demand forecasting methods. The term forecastability’ means the degree to which a product’s demand may be forecast. The best way to judge forecastability is by evaluating all data readily available. Some relevant examples of available data are marketing campaigns, historical sales and price figures, historical leftover volume, or stores and product directory.
Not all data are created equal, and attempting to implement a standard or general forecasting system can potentially lead to inaccurate forecasts. Product segmentation is one of the most important steps in preparing data, which includes growth brands, harvest brands, new products, and niche brands. This type of approach increases customer value by keeping the focus on top tier value products and allowing us to capitalize on available opportunities in the market.
This pandemic will end, although the timeline is not currently clear, and at that time demand will return to normal patterns and levels. However, the data gathered during this period will be of immense value. Incorporating that data into our future forecasting models, we will be more accurate in predicting future demand patterns in the face of similar crisis events.
This data we’ve collected during the COVID-19 pandemic will act as a training data set of sorts, allowing us to improve our forecasts during the next similar crises the world faces.
