Feature engineering is an art and science of converting raw data into relevant features. By carefully selecting and transforming raw data into meaningful features, you can highlight the most relevant patterns and relationships that influence demand. This process involves everything from identifying key variables like seasonal trends and economic indicators to creating new features that capture complex interactions within the data. Essentially, good feature engineering ensures that our AI/ML models have the best possible information to work with, leading to more accurate and insightful demand predictions.

For example, consider a new model year launch for an automobile. We can use the model year launch date and create multiple useful features from it, such as time since launch, time till next launch, etc.

Model training and forecasting are like the heart and soul of AI/ML-based demand forecasting. During this phase, the model learns to recognize patterns, trends, and relationships within the data, much like how we learn from our past experiences. Once the model is trained, forecasting comes into play, where the model uses its newfound knowledge to make predictions about future demand. By continuously refining and updating the model with new data, we ensure that our forecasts remain accurate and relevant in an ever-changing market landscape.

To make sure our forecasts are hitting the mark, we measure their accuracy using metrics such as weighted mean absolute percentage error (W-MAPE). W-MAPE helps us understand how far off our predictions are, while also giving more weight to periods with higher sales, so we’re not just averaging errors but focusing on what matters most to the business. On top of that, we use back testing, which is basically running our model on historical data to see how well it would have predicted past outcomes. This process helps us build confidence in the model’s reliability before we rely on it for future decisions. Together, these steps give us a clearer picture of how trustworthy our forecasts really are.

Driver decomposition aims to understand and predict the factors influencing future outcomes. It is a way that allows us to dig deeper into what really drives demand.

Lets look at an example (Illustration 1). We see that the demand forecast for the product is influenced by three primary drivers: price, promotions, and macroeconomic factors. The analysis shows that out of the 108 units forecasted, 100 units are attributable to historical demand (time series-based forecast) while 10 units are due to price adjustments, 5 units are due to promotional activities, and -7 units are due to macroeconomic factors. 

While driver decomposition helps break down the demand into relevant drivers, it’s important to understand the sensitivity to these drivers to address the most influential driver for demand shaping. Coefficients essentially quantify the impact of each driver on demand, showing how much a change in one factor, like price or promotion, will affect sales. Larger coefficient values indicate a stronger impact, while the sign (positive or negative) denotes the direction of the relationship. For instance, if the coefficient for a promotional campaign is high, it means that promotions significantly boost demand.

Let’s consider another example (Illustration 2) to illustrate this further. In this case, we’ll examine how the “price” driver affects the demand for different products sold by an organization. For Products E and F, we see a very low negative coefficient, which means that price isn’t a major factor influencing their demand. However, if the price does go up, we can expect a slight decrease in demand—like a gentle nudge rather than a shove. On the other hand, Product H has a high negative coefficient, indicating that price is a significant driver for this product. This means that any increase in the price of Product H would lead to a substantial drop in its demand—like pulling the rug out from under it!