The challenge of short series and new product launches
Artificial intelligence has emerged as an invaluable resource in the retail sector, providing a variety of advantages including demand forecasting, price optimization and personalized recommendations, all influenced by temporal requirements.
In the dynamic world of retail, the lack of solid historical data makes it difficult to accurately estimate future demand. This uncertainty about how many products should be produced or purchased to meet market needs poses a constant challenge for companies ranging from small retailers to e-commerce giants.
This challenge affects marketing and sales teams as well as supply chain management and logistics teams.
It can lead to problems of overinvestment in production, resulting in inventory surpluses and financial losses, or underinvestment, resulting in lost sales opportunities and customer dissatisfaction. In a highly competitive market, companies that can anticipate demand have a competitive advantage, while those that do not anticipate market trends risk falling behind their competitors.
As a result, companies are looking for innovative solutions to overcome these challenges. Artificial intelligence (AI) has emerged as an invaluable resource in the sector. Using machine learning algorithms and real-time data analysis, AI offers a variety of advantages in retail, not only in sales forecasting, but also in other common problems such as price optimization or personalized recommendations.
However, what happens when we launch a newly created product? What happens when we launch mini-series of seasons in fashion? There is no historical data here for the systems to learn from, and we must look for alternatives.
Five proven methods for non-historical sales forecasting
When faced with modeling time series with limited historical data, we use creative strategies. We must differentiate between two scenarios:
- If we have limited data in our time series compared to others, we must try to find correlations between the available data and more complete data sets.
- If the time series we want to predict is new, but contains attributes such as color, brand name or product category, it is possible to segment these attributes to identify similar products with available history, and use that information to make predictions about those that do not.
Even if we have a combination of temporal data and comparative attributes, we will take advantage of both to improve our predictions.
In the following, I list a number of methods based on examples to try to understand it better. These methods are not independent and can be combined with each other in order to improve predictions.
Clustering to identify similar products
Suppose we launch a new type of green running shoe. Through clustering, we detect that it is grouped together with other high-performance running shoes aimed at professional runners, but also shares characteristics with eco-friendly products in other categories. This analysis allows us to look at sales trends and consumer behavior in these two segments to infer the potential of our product.
The clustering technique is essential for grouping products based on similar characteristics, especially useful when direct historical data is not available.
2. Use of initial data for adjustment
If our new smart refrigerator sells 500 units in the first month when the model had projected 800, it is good to adjust the model to gauge expectations and better understand the factors behind the discrepancy.
When the first sales data is available, it can be used to refine initial predictions by adjusting the projected scale or trend.
3. Hierarchical models for capturing interdependencies
In the fashion industry, if a summer collection gets an exceptionally positive reception, it is likely that all products in that collection will experience an increase in demand. A hierarchical model would allow adjusting the predictions for each item in the collection based on the overall performance of the collection.
In sectors where products are launched in collections or have strong interdependencies, hierarchical models can be hierarchical models are particularly useful.
4. Predictions for functionally equivalent products
Let’s look here at the online sale of a yoga pant. Instead of focusing on predicting sales of a single yoga pant model, we evaluate the performance of the entire yoga pant category over time, which gives us a more accurate understanding of the overall demand trends in that product category.
Aggregate category forecasts can be extrapolated to guide new product launches within the same category.
5. White-box models and the use of explicit causal relationships
When launching a replacement part for a specific car model, we develop a model that takes into account the number of vehicles of that model still in use, the average life of the component and our aftermarket penetration. If we know that there are 100,000 cars of that model in circulation and the component tends to fail every 5 years on average, we can make a reasonable estimate of the annual demand, adjusted by our expected market share. For products for which the demand can be explained by known factors, the white-box models that incorporate explicit mathematical relationships are extremely valuable.
Each of these approaches not only helps us to predict demand for a new product but also to better understand the market and how our product is positioned within it. The key is to iterate and constantly adjust the models as new information becomes available, allowing for a more informed and dynamic market strategy.
Best practices learned from real projects
Finally, I would like to share a couple of concrete examples of real projects in which we have implemented these techniques with beneficial results for our clients.
For a multinational company in the food industry, we used clustering algorithms to segment products according to their attributes. From these segments, we evaluated multiple approaches in collaboration with the client. We detected that demand was strongly influenced by local and cultural factors, such as the region of sale. This led us to select the strategy most aligned with the operational reality of the business.
For example, we observed that demand varied significantly by province, with correlated sales patterns within the same region.
Never forget to validate the possible influence of external factors, which will significantly improve the accuracy of your sales prediction model.
Let us now take the example of a fashion industry. This particular challenge was related to the interdependence between products in this industry. A client in the fashion industry needed to forecast the sales of its new collection, but faced the problem that the performance of each item was influenced by the overall success of the collection.
Using our predictive models, we were able to capture this “collection” effect effectively. We implemented hierarchical models that allowed us to analyze how the performance of each item was correlated with the overall performance of the collection. As a result, we provided the client with a more accurate and complete prediction of the sales of their new product line, allowing them to make more informed decisions about inventory management and marketing strategies.
Finally, in our use cases, we understand that these solutions demand close collaboration between data scientists and business professionals. It is crucial to identify and analyze meaningful relationships between time series, as well as to validate their consistency and relevance from a business perspective. Therefore, time series analysis is not limited to simply applying sophisticated algorithms, but also involves understanding the business context and rigorously validating predictions.
