Every inventory manager lives with the same dilemma. Excess inventory freezes capital and fills the warehouse with goods that may become outdated or expire; insufficient inventory leads to stockouts, lost sales, and disappointed customers. This is the classic newsvendor problem, and for decades the toolkit for solving it was statistical in nature. In recent years, machine learning has ceased to be a buzzword in this field and has begun to prove itself over classical demand forecasting models.

The Models We Learned Still Work, But Under Certain Conditions

There is no point in dismissing the veteran tools. Moving average, exponential smoothing, and ARIMA models still provide reasonable demand forecasting; the EOQ formula determines the economic order quantity; and safety stock is calculated as the product of the desired service level, the standard deviation of demand, and the square root of lead time. These models work well when demand is relatively stable and consistent. A significant portion of organizations in Israel manage inventory exactly this way, mostly within the ERP system, and for a large portion of items this is completely sufficient.

Where Classical Models Don’t Deliver Results

The problem begins when the forecasting needs to become more dynamic. Commodities, weather, promotions, local disruptions, suppliers changing prices and long tail items — all of these break the assumptions of the linear model. Above all, there is the Bullwhip effect, a small fluctuation in consumer demand that amplifies through the supply chain. Walmart learned how its linear methods, derived from historical patterns, failed to stand up to a sudden product trend, couldn’t meet the demand of rapidly changing items, and didn’t identify that their products were becoming obsolete.

What Machine Learning Brings

The advantage of the new models is their ability to simultaneously consume dozens of external variables: sales history, price, promotions, weather, special events, and even social media signals. Algorithms like Random Forest and XGBoost, and neural networks from the LSTM and GRU family, capture long-term non-linear patterns that classical models struggle with. Research from 2025 showed that XGBoost achieved weekly sales accuracy at Walmart with an R² of 0.97, meaning the model explains 97% of demand variance. The main findings are that the model is more accurate at identifying future trends, not just extrapolating from the past. In recent years, the combination of deep learning and time series modeling is yielding breakthrough results in inventory prediction.

Field Examples

Walmart is the boldest example. With millions of items and thousands of stores, the company, an e-commerce platform based on deep learning, automatically factors in weather, local events, and finally uses GPT-4 to improve inventory forecasting. The supply chain optimization department documented this in the INFORMS journal in 2024. The result: fewer stockouts, lower inventory levels, better availability. Amazon and Zara, as documented in the 2024 industry reports, went further in their approaches. Amazon developed deep learning-based forecasting with natural language processing capabilities to read social media trends and predict demand for new products. Zara, on the other hand, uses machine learning models driven by the store-level sales data it collects twice a week to minimize overstock and stock-outs, recording about 8% annual profit growth according to IHL.

The 2024-2025 Frontier: Foundation Models for Time Series

The most exciting recent development is foundation models for time series, such as Chronos and TimesFM. These models, trained on massive amounts of data, are able to predict a new series as if it had almost no entry barrier for organizations that have no large volumes of training data. The field is still young, but it shows the same trajectory.

Caveats

Machine learning is not a magic wand. It requires clean data, sufficient volumes, and the human skills to maintain it, and 41% of organizations report difficulty running it due to lack of proficiency. For completely new items or very slow-moving items, it has no history to learn from, the statistical methods and human judgment still produce better results. Beyond that, the model can explain the logic of the decision, meaning the algorithm’s business decisions that need to fit the context: there is no one-size-fits-all.

The Model Is Only Half the Story

Algorithms are not a substitute for inventory policy. The real value comes from combining smart inventory management with smart data policies: ABC and XYZ segmentation, service level assessments per item, inventory replenishment structure and S&OP process. This is not just a technology, not just a tool — not all partners are the same. In the supply chain department, Mashik combines both worlds, the methodology and the technology, so the improvements are implemented in practice for optimal inventory management and improving service levels to the customer.

Main Sources

INFORMS Journal on Applied Analytics 2024; World Journal of Advanced Research and Reviews 2024; IHL Group; XGBoost/LSTM sales forecasting research 2024-2025; Foundation models for time series (Chronos, TimesFM) 2024-2025.