Advanced Demand Forecasting Part 3: Need-to-know’s when considering ADF

Advanced Demand Forecasting Part 3: Need-to-know’s when considering ADF

Data model analysis and improvements
Data model analysis and improvements

Advanced Demand Forecasting is the next-level solution for businesses looking to identify and understand which factors have the most significant contribution to demand so that they may invest with greater accuracy. Demand frequently does not have a simple, straightforward relationship with any singular feature that organizations can increase to drive demand. That’s where Advanced Demand Forecasting steps in.

Advanced Demand Forecasting deploys feature engineering to enable organizations to recognize those inter-woven relationships and recommend when and how to adjust variables with the greatest impact on demand. Let’s use an example scenario to show one such inter-woven relationship between features and how reality often presents both internal and external influences.

Example use case

To provide an example, consider the fictional retail chain, Retail X. Retail X may have multiple stores, but we will look at this example on an individual store level instead of the whole company.

In this example, Advanced Demand Forecasting determined Retail X’s sales total is influenced by several factors, one of which is the total number of Sales Associate hours. At the example store, our model shows decreasing sales primarily due to the number of Associate hours.

At this point, other demand forecasting models may indicate that one Sales Associate hour is worth $100 in sales at the store, resulting in increased Associate hours due to an assumption that increasing hours translates directly to increased sales.

While Advanced Demand Forecasting will indicate a relationship between Sales Associate hours and sales, it can take these insights a step further by showing their relation to other demand factors. For example, it may indicate that investing more person-hours or employees into a store doesn’t increase foot traffic and may result in diminished returns on growing sales. By understanding this nuance and the relationship between staffing needs and foot traffic, Retail X can drill down to determine the issue isn’t store staffing, rather increased wait times during peak hours. This kind of insight allows them to act with greater precision to maximize sales with minimal increase to overhead (e.g., unnecessary sales staff during slow periods).

At this point, Retail X can conclude that increasing employees by N amount during peak hours will help reduce wait times, resulting in increased sales. In the end, Advanced Demand Forecasting illustrated that Associate hours were related to two other factors, highlighting the importance of considering all elements that help generate sales.

Importance of clean data

Now that you understand how features will often influence one another and demand is not dependent on any single variable, it should be apparent that having useful data is imperative! Demand regularly fluctuates, especially in the retail and consumer goods industries. To forecast as accurately as possible, here are four key things to remember for Advanced Demand Forecasting to have maximum effectiveness:

  1. How far into the future Advanced Demand Forecasting can predict depends on how stable the demand is. Businesses with higher demand volatility will produce shorter forecast windows than demand that is more stable year-over-year.
  2. Predictions are only as good as the data on which they’re based. Advanced Demand Forecasting relies heavily on internal data, as machine learning requires historical data to develop accurate forecasts.
  3. Raw data and raw counts are best. Predictions are more accurate when the model can learn from raw or unaggregated data from previous years. Aggregated data will only tell a portion of the whole story or show a percentage of a factor (such as X% of people redeemed a given coupon) and may not produce meaningful results.
  4. Daily operational data at the individual store level is necessary. Knowing information like promotion usage rates, foot traffic volumes, operation hours, staff information, product information, etc., can help contribute to an accurate forecast.

These four points apply to virtually any demand forecasting model. If you have any black boxes or inefficiencies within your demand models, first check to see if the data you’ve been feeding your models are clean and organized. Remember, the quality of your data ties directly to the quality of your modeling.

If your data is clean and organized, then there is a strong likelihood that your current models have experienced concept drift or that your forecasting solution falls short of your expectations. Either way, we can help! Contact us to learn more.