Harnessing the potential of computer vision and AI
Generative AI (GenAI) presents a significant opportunity for fashion businesses to enhance efficiency and customer satisfaction. While the industry has already begun to explore basic AI applications, GenAI capabilities remain largely untapped. Despite this, notable use cases have emerged, particularly in areas like demand forecasting.
This technology leverages unstructured data such as text, images, and video to create new forms of media, including 3D designs and extract product attributes. With retailers typically planning designs months in advance based on guesswork due to a lack of historical transaction data, GenAI offers a welcome solution, providing an unprecedented edge in the ever-competitive marketplace.
AI-forward into forecast precision
In the dynamic world of fashion retailing, one persistent challenge is determining precisely what and how much to buy. This issue stems mainly from the difficulty in predicting demand, especially for new products where historical data may not exist. We can use data science and automation to boost forecast accuracy throughout the critical stages of the forecasting process in the following ways:
• Use comprehensive internal and external data: Improve demand forecasting accuracy by applying data-driven methods. This involves analyzing internal data like sales, orders, inventory, and promotions and external data like events, competitor actions, and weather. For in-season forecasting, to manage inventory and open to buy, factors like current stock and incoming orders help understand what customers might buy soon and how much more to buy.
● Data cleaning: Detect and manage outliers using algorithms and fill in missing data.
● Feature engineering: Automatically spot demand patterns for numerous products by creating new features that capture seasonal trends and demand changes.
● Synthetic data creation for new products: Overcome the challenge of forecasting new products by using automation to find similar products from existing data and generate synthetic data, then switch to actual historical data after a certain period.
● Ensemble approach and automated algorithm selection: No algorithm provides the best result for fashion skus. Customize forecasting techniques to suit the unique characteristics of different product categories, segments, and channels. For example, develop specialized methods for forecasting new products without historical data. Boost forecast accuracy using various forecasting algorithms such as ARIMA, SRIMAX, Neural Network, Halt and Winter, Custom regression including ensemble methods, and create automated systems to pick the best algorithm for each product category or item.
● Machine learning algorithms for causal variables: Employ sophisticated machine learning algorithms to identify the impact of causal variables and interactions among multiple products.
● Customize the test and train data range: We can customize the data range to build the algorithm and train it according to data availability, product characteristics, and forecast purposes.
● Loop in human intelligence and business context: Employ fashion business understanding, such as the reason for excess inventory at the beginning of the season, any delay in time to market for new products, etc., to boost accuracy.
With this in mind, fashion executives should prioritize building tools that offer tangible value rather than unquestioningly experimenting with existing ones. By applying the principles above, Fractal has significantly reduced the forecast error for a leading apparel and footwear brand.
Demand forecasting is a very important component of our retail planning process. A high-quality forecast helps the business reduce overstock/understock through efficient inventory replenishments and bring valuable insights toward long-term sales trends. Fractal helps establish science-based, state-of-the-art statistical forecasting models for granular levels, including new product sales forecasting. [We] really appreciate the hard and intelligent work provided by Fractal.”
-Snr. Director, Fractal client
Key fashion business challenges solved by AI Foundation models
1. Seasonality
AI foundation models can help us tackle specific challenges in the fashion industry, like predicting sales patterns influenced by seasons. While basic models can spot weekday versus weekend sales trends, machine learning goes further. It considers things like promotions, discounts, and even the weather to give a more precise forecast. This comprehensive approach gives a complete picture of what to expect in sales in the upcoming days, weeks, or months.
2. External factors
Forecasting demand is challenging, especially when considering the weather, local events, and what your competitors are up to. AI foundation models makes this job easier by quickly figuring out how these factors affect sales.For example, it can tell us which products people are likely to buy more when it’s raining or if a big event is nearby. It can also highlight products that might not sell well in certain situations. Plus, machine learning can show us how competitor promotions or new stores might affect how many people come into our stores or visit our website.
3. Pricing, assortment, and marketing changes
Changes such as retailer offers, promotions, display alterations, and marketing investments can influence product demand, but predicting their impact requires complex analysis. AI foundation models can handle this task and accurately forecast how adjustments in pricing, product selection, and marketing efforts can affect demand.
4. Long-tail products
Long-tail products often have very few monthly sales, so more data must be available to analyze, leading to unpredictable demand patterns. The predictions can become unreliable when you add in other factors like external influences and different sales channels. However, advanced algorithms can filter out this noise, prevent overly specific predictions, and still give accurate results even for niche products.
5. Omnichannel
Forecasting for stores and online sales presents different challenges, with variations even within each store and channel due to factors like location and logistics. AI foundation models allow retailers to analyze data from all stores and channels, understanding the unique needs of each. This enables suggestions for stock movements within stores and across channels, like transferring excess in-store stock to online inventory and vice versa, optimizing inventory management.
6. Unstructured data
AI foundation models are best suited to handle unstructured data. They allow retailers to process data in real-time. Inputs such as search trends, social media actions and hashtags, global and local news, and other non-linear and unstructured data help machine learning algorithms increase the accuracy and value of their output.
The impact of Generative AI and attribute-based forecasting
In contrast to traditional SKU-level data forecasting, attribute-based forecasting leverages advanced analytics to predict future trends based on product attributes. These attributes encompass factors like color, size, brand, price point, or any other feature that characterizes a product.
Traditional demand forecasting often struggles with ‘long-tail,’ niche, and new items — products with low demand, infrequent sales, and newly launched. The scarcity of historical data complicates accurate predictions for these items. However, attribute-based forecasting directly addresses this issue. Emphasizing product attributes generates reliable forecasts even for such products.
One common challenge in attribute-based demand forecasting is the unavailability of attributes in master data, specifically for new products before they are launched in the market. Making apparel product attributes data involves categorizing, defining, and linking attributes to products while ensuring accuracy and consistency amidst the ever-changing retail landscape. Challenges include maintaining data accuracy and consistency, handling attribute differences, and staying current with trends and customer preferences.
Generative AI collaborates with AI foundation models, forming a powerful blend of predictive capability and creativity. This synergy enables fashion retailers to effectively predict demand based on attributes.
Attribute extraction using computer vision and NLP
Having accurate and detailed information is essential with a wide range of products. To tackle this, we suggest a new approach: using Computer Vision and Natural Language Processing (NLP) to automatically gather detailed attributes from images and text of the products through the below steps:
Data collection and preprocessing:
● Gather high-quality images of apparel products from the retailer’s inventory.
● Label images with attributes like color, pattern, material, and style.
● Collect product descriptions, specifications, and other text information.
Computer Vision for image analysis:
● Use Convolutional Neural Networks (CNNs) to analyze product images.
● Develop models to recognize visual attributes such as color, pattern, material, and style.
● Extract relevant data from images and convert visual features into organized information.
Natural Language Processing (NLP) for text analysis:
● Create NLP models to analyze product descriptions and specs.
● Implement Named Entity Recognition (NER) to identify and extract attributes from text.
● Process and standardize textual data to ensure accurate attribute extraction.
Attribute synthesis and enrichment:
● Merge information from images and text to build complete attribute profiles for each product.
Tech-forward fashion
In the not-too-distant future, the fashion industry’s narrative will no longer be about chasing trends or playing catch-up with consumer desires. Instead, through the robust capabilities of GenAI and AI foundation models, it will be about orchestration — seamlessly aligning consumer desires, production capabilities, and environmental considerations.
As the digital revolution unfolds, embracing these technological advancements will become necessary, not a choice. Investment in these technologies today can yield significant returns. Improved demand forecasting can reduce inventory carrying costs and markdowns and, thus, increase full-price sell-through and customer satisfaction with better product availability. Thus, overall, it will reduce operational costs and increase sales and profitability. Holistic demand management, in this scenario, will be more than just an operational strategy. It will be the ethos around which the fashion industry revolves.