Meet LDSM, Syrup’s Neural Network-Powered Demand Forecasting Model

When it comes to boundary-pushing AI technology, nothing piques my interest more than neural networks (NNs). They are the foundation for large language models (LLMs), which have made headlines for their applications in generative AI.

By scraping the internet, these incredibly sophisticated and complex models can do some truly mind-blowing things: create art from a simple prompt, design new genes, find anomalies in the stars…

But retailers and brands sit on a treasure chest of data that is even more powerful for them than “just” the internet.

Start with the rich history of point-of-sale transactions and daily inventory snapshots buried in an ERP. And then add e-commerce data hidden in web analytics reports; all the product information that lives in a PLM tool; customer review data floating on the web; advertising spend and other promotional activities.

A fortune in data, but a headache to analyze. Until now.

I’m incredibly proud to announce the release of the Large Demand Sensing Model, or LDSM for short.

Leveraging the power of cutting-edge neural networks in combination with our proven and sophisticated "traditional" predictive AI algorithms, LDSM allows retailers to feed all of this rich internal data into an intelligent brain — meanwhile, that brain consumes relevant external inputs like weather forecasts, local events databases, and social media trends.

The end result? Unbelievably precise demand predictions constructed from a dizzying maze of unstructured data inputs.

We’ve built LDSM from scratch to meet the unique needs of apparel and footwear businesses, training it on industry-specific proprietary data.

Rest assured, all of that forecasting insight is not wasted by dumping it into an Excel sheet. Our Platform, and the system of models that underpin it, power all of our elegant Allocations and Buying workflows. Each is designed to help omnichannel retailers take predictive inventory actions that drive profitability.

“By combining state-of-the-art AI models with proven and sophisticated algorithms, Syrup's technology enables retailers to harvest both their data and external data in order to improve their customer experience through margin-accretive recommendations.”
‍
Dr. Thomas Vetter
Former Senior Vice President and Global Head of Product Management, Consumer Industries at SAP

Why Neural Network Forecasting is a Game Changer for AI-Driven Inventory Optimization in Retail

For the uninitiated, neural networks are a type of AI model. Here’s a quick primer on how neural networks work if you’d like to catch up.

While traditional machine learning algorithms (for example, linear regression or decision trees) have a relatively simple structure, artificial neural networks are complex and intertwined — much like the human brain. This sophistication supports more accurate and granular predictions, even with the dynamic demand and supply inputs that are standard for apparel and footwear brands.

Neural network models’ flexibility allows planners and analysts to uncover intricate relationships between their inventory characteristics, omnichannel networks, and pricing approaches. With surgical precision, the model can zero in on how granular components like SKU-store or SKU-node combinations contribute to demand changes — and with equal ease, it can zoom out for a more holistic analysis of category- or region-level interactions. This level of detail is crucial for retail demand forecasting, especially in AI-driven inventory optimization strategies.

As with most AI models, neural networks learn over time. But unlike other models, neural networks benefit from the ability to apply embeddings, a unique tool increasing the granularity of relationship learned. They don’t stop at understanding how your demand changes over time — they learn how individual SKUs, stores, and other factors interact over time. Increased specificity provides clarity on which demand levers will drive the greatest impact.

What’s especially cool about this is how different altitudes of analysis can be combined to form an ultra-powerful tool. We can train LDSM to solve particular aspects of the demand funnel — for instance, a component to learn trend, a component to learn seasonality, a component to learn correlations across SKUs/stores, and a component to learn holidays/promotional effects. The network can reconcile across all of these components to learn total demand at that SKU-store level.

Neural networks are also much more efficient than other model types when we need to apply learnings from previous embeddings into new contexts. For example, if you’re opening a new store and want to predict how an existing item will perform, the model can quickly extract existing SKU-store signals and transfer it to new SKU-store combinations. Another example: if you’re introducing a new style the world has never seen before (also known as the “cold start” problem), the model can quickly extract signal from your entire catalog of data and transfer it to this net new product. Other models struggle with this level of feature modeling.

And when it comes down to it, neural networks just perform better on average than many model alternatives on key metrics like bias and accuracy.

“One of the biggest value drivers is that specificity. It’s the difference between swinging a hammer at your problem and carving it with a laser.”
Mike Smith, PhD
Senior Manager Data Science at Syrup

The Impact of Neural Network Demand Forecasting on Omnichannel Merchandising

Of course, all that predictive power would mean little without a way to make use of it. LDSM is just one of many AI models we utilize to support demand forecasting and inventory optimization. This suite of models functions as our Platform’s engine, powering each of our apparel and footwear merchandising workflows.

Building on our existing functionality, which we have already used to help major omnichannel brands and retailers like Salomon, Desigual, and Reformation reduce carrying costs and optimize sell-through rates, access to the power of neural networks unlocks some brand-defining new opportunities.

Here is a small sample of the merchandising tactics neural network forecasts can support:

• Identifying viral events and anticipating region- or even store-specific demand shifts so you can quickly react with inventory rebalancing or just-in-time purchase orders
• Confidently launching brand new products with the ideal inventory set to maximize full-price sell-through without waste
• Anticipating cannibalization effects of new product introductions
• Elevating assortment or merchandise planning activities with unimagined specificity in prediction

As we continue to add new workflows to our product suite, LDSM will remain ready to support them, where appropriate — that’s the value of building on a modern SaaS infrastructure designed for constant iteration.

“Neural networks and LLM-like models are the next frontier to unlock exponential value in merchandising and planning. LDSM provides cutting-edge AI advancements that any merchandising team can easily leverage.”
John Andrews
Former CEO, Celect (acq. Nike)

What’s Next?

Having a proprietary neural network model built by our in-house data scientists is not enough for this hungry team. With our architecture in production, we can continue to evaluate and develop new features and capabilities:

• Continuous iteration and improvement on model embeddings
• Hierarchical modeling directly in the network
• Leveraging sequential data as features
• Customizable loss functions
• Sandwiching multiple neural network models focused on specific data relationships

And if that means little to you, don’t worry. The key takeaway is that we are not done innovating — and every improvement we make stems from one question: is this going to improve business outcomes for our customers?

Be among the first to experience the future of demand forecasting. Request a demo today and start seeing results.