How Santa is using demand forecasting to plan the timely delivery of gifts

By Team Ikigai

How Santa's using demand forecasting to plan timely gift delivery

The Christmas holiday season is well underway and is the busiest time for supply chain managers. It takes months of planning and preparation to ensure that everything leading up to the big day goes smoothly. Unfortunately, not everything goes as planned.

Unlike major companies managed by supply chain professionals, Santa Claus is one individual who has never had a riot-like situation. Not only can he know when you are asleep to deliver at the right time, but he can correctly predict demand to guarantee that everyone receives the Christmas gifts they wish (at least with an impressive hit rate). For this reason, Mr. Claus is regarded as an industry leader in supply chain management.

Santa Claus's stories inspired us to consider what lessons we can learn from them.

Planning is important

We all remember that one youngster from grade school who never stopped complaining because Santa Claus did not bring them exactly what they wanted. They created the impression that it was the end of the world.

While Santa never likes seeing sad children on Christmas Day, his elves (or the child's wise parents) occasionally make mistakes. However, these are typically minor, and many youngsters receive their desired Christmas gifts due to thoughtful demand planning.

Santa is not the only person who must meticulously plan for success. Demand planning is an integral component of every supply chain. No supply chain management wants to be left with empty shelves because there is insufficient inventory to fulfill demand. Nonetheless, supply chain managers want to avoid stocking shelves with products that no one wants to buy.

Balancing this equation takes considerable work, especially if you attempt to do so manually. Using only historical data can frequently result in disparities between forecasted and actual demand due to omitting new elements as outliers. Therefore, many supply chain managers will utilize clever algorithm-based forecasting solutions.

Demand forecasting is the time-tested practice of predicting customer demand over a specific period using previous data, traditionally on purchases. Almost every industry employs demand forecasting to optimize food distribution on store shelves, hardware compute-power in a data center, or (hypothetically) placing gifts under a Christmas tree brought by a cheerful elf.

The most accurate results can be reached by combining Artificial Intelligence (AI) and Machine Learning (ML) with a demand planner's practical knowledge and experience. Having more accurate forecasts helps all the functions downstream in the supply chain, such as manufacturing and procurement, and increases earnings and customer satisfaction. This is Santa's objective and should be the guiding principle of every supply chain.

Building Accurate Demand Forecasting Models

In general, there are two types of ML models for demand forecasting:

  1. time series models and
  2. regression models

Forecasting sales based on historical data is possible using time series models. Time series models typically do not require features — only actuals, such as previous sales data on socks, to predict future sales — and hence do not risk feature drift.

A disadvantage of time series models is that they frequently require years' worth of data to generate good forecasts, limiting their utility. For instance, a consumer electronics company may debut a new category of wearable devices or a superior smartphone, which lacks comparable historical sales data because they are new product categories or inventions.

Even when such data are available, the overall accuracy of time series models based only on historical data may be lower than that of models based on characteristics. And even though drift is less of a problem, performance still worsens in the real world since previous data frequently do not convey information about future events (i.e., COVID-19 contributing to a dip in the stock market in 2020).

Regression forecasts can use more complex models to provide more accurate predictions when used to predict the quantity demanded of a defined period (e.g., ten days in the future) or the quantity demanded "n" day in the future, with the number of days (n) as a feature. In contrast, regression forecast models do not require the same level of historical data. Because the same schema may be used for additional segments, they are also reasonably simple to upgrade and retrain.

For these reasons, regression forecast models are highly prevalent across businesses, informing anything from how airlines hedge fuel price surges to how retailers price holiday end-caps and other high-traffic items. Complexity and the balance between bias and variation can make these models more susceptible to drift.

This is not necessarily an either-or choice. Numerous organizations find that regression and time-series models are helpful in a variety of circumstances. Some companies monitor a time series forecast and a regression forecast alongside sales data (actuals), segmenting models further by factors like production, location, or season.

Highlighting the drifts: Model Observability and Model Monitoring in Demand Forecasting

As with most predictive modeling problems involving future occurrences, demand forecasting is regarded as academically and practically challenging due to the numerous unaccounted-for uncertainties in a model when a prediction is made.

Businesses must handle changing consumer preferences, unprecedented demand, and a more complex supply chain characterized by inflation, delays, and other unforeseen issues. 44% of CFOs believe that delays and shortages across the supply chain increase expenses, with 32% reporting falling sales.

Since the results of demand forecasting models are frequently employed in planning, the cascading effects of degraded performance may take time to determine. However, each instance of performance degradation, model, or concept drift might result in significant financial losses.

Like Santa's elves would know and tell him when things aren't going right, AI/ML monitoring and observability are essential for notifying teams when these events occur, evaluating the size of their impact on models, and gaining insights into the fundamental causes of problems so that they may be resolved swiftly. An ML observability plan could be the difference between a shop having sufficient inventory to satisfy Christmas demand and losing millions of dollars in sales due to out-of-stock products.

Measuring Demand Planning Performance

  1. Mean Error (ME): the average historical error (bias); a positive value indicates overprediction, whereas a negative value suggests underprediction. Although the mean error is not often the loss function that models optimize for during training, the fact that it monitors bias makes it a valuable metric for assessing business effect.

  1. Mean absolute error (MAE): the arithmetic means of the absolute value difference between a model's predictions and the ground reality throughout the whole dataset. Excellent "initial impression" of model performance, as the high errors of a few forecasts do not bias it.

  1. Mean absolute percentage error (MAPE): estimates the average magnitude of error produced by a model; one of the most prevalent metrics for assessing the accuracy of model predictions.

  1. Mean squared error (MSE): the squared and averaged over the dataset difference between the model's predictions and the ground truth. MSE is utilized to determine how well-predicted values match actual values. As with root mean square error (RMSE), this metric assigns a greater weight to large errors and may be helpful when a company wishes to penalize outliers or errors of significant size aggressively.

It should be highlighted that mean error is insufficient for describing biases. Mean error can be canceled out by opposing metric values during a feature drift event characterized by both over-prediction and under-prediction during a specific period. Because of this, it is advantageous to compare the magnitude and direction of mean error and mean absolute error — together, they can be used to determine when a model's performance over a certain period is unsatisfactory.

If, for instance, a big shop has a high mean absolute error but a negative mean error, it is likely underestimating demand. The retailer's ML team may learn this is due to shifting consumer preferences, such as purchasing holiday gifts earlier and in larger sizes than in previous years. Armed with this knowledge, they can suitably retrain the model. With an ML observability platform, firms can instantly view and identify the core cause of problems. This becomes considerably more crucial during outlier events.

Timing is essential

It would be an understatement to suggest that this would not be fantastic. Imagine waking up on Christmas morning to find no gifts beneath the tree. If you have children, you know it would be an absolute disaster.

Timing is crucial for Santa Claus. Along with accurate forecasting, he must ensure everyone's Christmas gifts are delivered by December 25. This is also essential for logistics businesses to ensure Christmas Day delivery of gifts. Perfect punctuality is desirable throughout all aspects of the supply chain, but it is especially critical in the logistics process.

In a world where e-commerce sales are increasing, with more and more people shopping online, there's more focus on shipping and delivery alternatives.

According to a survey by PwC, roughly a quarter of respondents indicated that speedy delivery would influence their shop choice. However, consumers' conceptions of "rapid delivery" are evolving. If this trend continues, consumers will be more demanding and less forgiving when delivery takes longer than a day. This unprecedented imprecision results from the increased availability of next-day and same-day shipping.

Closing Reflections

Whether you believe in Christmas or not, you cannot deny that the Christmas supply chain works for most people. This is miraculous. Santa's forecasting and logistics models are a fantastic example to follow if you strive for perfection in your supply chain.

Given the changing nature of product development and global operations across industries in the wake of COVID-19, demand forecasting models' applications may undergo a once-in-a-generation reset. Retailers, for example, are likely more ready than in previous years to take on excess inventory, as they consider under-forecasting more costly than over-forecasting due to the risk of losing consumers.

AI/ML can assist teams in optimizing for these outcomes, avoiding costly errors, and staying on top of rapid changes. By identifying the factors contributing to over- or under-predictions and rapidly resolving concerns, teams can ensure that future forecasts are accurate and positive. Here is an example of how an ML observability platform assists clients with demand forecasting.

Ikigai's operational BI platform transforms the way businesses make tactical decisions. Business-user-friendly UI/UX enables anyone to infuse and prepare data and run robust AI-powered analyses to achieve their business goals.

Ikigai combines data analytics, visualizations, and automation with its proprietary technologies, such as DeepMatch and DeepCasting time series forecasting, and translates them into more precise decision-making data. Let's connect for a quick chat.

Best wishes for Christmas!