It is no secret that any company dealing in goods lives and dies by its profit margins and service quality. These variables intersect most strongly while maintaining inventory. A large inventory buffer is ideal for quick order fulfillment and product availability. However, it also leads to a higher cost and a higher risk of accumulation of unsold inventory, which will have downstream effects on warehouse availability and future order picking times. A razor-thin inventory buffer is ideal from a profit perspective as it will prevent accumulation and keep costs lower. But it will lead to product unavailability sooner or later.
Seasonal Fluctuations and Forecasting
These are the key considerations to keep in mind while optimizing inventory. This is further complicated by seasonal preferences for certain SKUs, rise in orders related to holidays and festivals, and effects related to marketing. It is easy to see how this can make inventory management a rather complex multivariate problem. Despite this, many organizations rely on institutional knowledge and relatively rudimentary methods such as monthly moving averages and linear regression to forecast demand and manage inventory.
These methods are often incapable of exploiting features built atop the effects described above and tend to rely on seasonal patterns and past sales. This can often lead to incorrect estimates. The next question therefore is: how can we make these forecasts better?
Machine Learning and Feature Engineering
The simplest answer is machine learning since it can model relationships between these features and exploit them to generate forecasts. It would be easy to say that we can throw everything in a neural network and call it a day, and some articles do indicate that they perform reasonably well. However, in use-cases such as this, explainability has a very outsized utility for the customer. Therefore, it makes sense to consider feature engineering and simpler models.
Feature engineering will often require a fair bit of experimentation based on the domain of the customer, and there is a strong argument to be made here about subject matter expertise. As far as modeling goes, the challenge there is to consider the time-series elements and add features dealing with seasonality. What those features might be, is out of the scope of this article.
Lastly, we must consider what explainability is. In its simplest definition, explainability is a method that tells us how a given set of data points resulted in a specific prediction. There have been a few methods proposed for it, but the one we choose is the well-defined method of using Shapley values. It gives us a clear answer about how a given result was obtained. This can tell us which features are strong and how they are affecting results that have great utility.
Schedule a demo with Hopstack’s product specialist to learn how the platform’s machine learning and forecasting features can help your organization adapt to fluctuations in demand.