Machine learning to predict demand

Demand for clinical services continues to grow. Forecasting this growth is complex due to significant fluctuations in demand driven by many factors including Black Swan events like COVID, seasonal demand driven by influenza, vacations, the start of school, weather and more. These causal factors are difficult to account for through the basic forecasting techniques employed by many laboratory services companies.
Predicting demand used to be all art—what one believed would sell. This non-scientific approach often led to either too little or too much inventory. When your inventory is highly perishable and expensive, like the consumables used in today’s modern laboratories, the impact to the business can be significant.

Machine learning and artificial intelligence are changing the world around us, both personally and professionally. We can now take in vast quantities of seemingly unrelated data and make sense of it. For example, we can look at the many factors that impact demand for a product to create a highly accurate picture of what demand will be. The application of machine learning has turned the once impossible into the possible.

Here we will explore three benefits to machine learning for demand forecasting that you can realistically achieve.

Accuracy and transparency

Stakeholders who care about forecasting in demand planning will also care about accuracy, and usually will not accept a new forecasting method unless it is rigorously validated against known forecasting benchmarks with proven accuracy. Machine learning for demand forecasting is highly accurate; this is proven over and over again in Kaggle competitions and modelling benchmarking studies.

For the more curious data scientist, machine learning for demand forecasting also has stable accuracy / bias trade-offs that can be adjusted on an ’efficient frontier’ of data science workflow, so that an accurate machine learning forecasting solution can be implemented quickly, and then studied over time to further improve the forecast. Furthermore, machine learning forecasting is not black box; the influence of model inputs can be weighed and understood so that the forecast is intuitive and transparent.

Greed for more data

An important contributor to accuracy is the ability of machine learning to ingest disparate data and leverage that information at a granular level to improve SKU-level forecasting (stock keeping unit). Simply stated, if data can be matched to the SKU at the point of sale or the point of distribution, the data can be leveraged with machine learning forecasting. Here, it is important to distinguish between predictive algorithms and the pre-processing of data that feeds them. Forecasting data, at its simplest level, has four data columns: Case ID; Time Series Member (like SKU at point of sale, which is the most granular level of forecasting); date of transaction; and transaction amount (in units or currency or volume) or transaction event (if the forecast focuses on events rather than amounts).

By ingesting this information, one can quickly build highly accurate and highly granular forecasts. All other available data, such as prices and discounts, distribution networks, weather information, social media ‘voice of the customer’ and advertising impressions that can be correlated with the data at a SKU or location or date level can be blended into the modelling database. These extra predictors are often important drivers of improved forecast accuracy and bias reduction, and machine learning forecasting incorporates these disparate data without manual data exploration and human intervention in the mathematical forecasting process. Again, this is an example of a greedy modelling process because a local optimal solution is attained for every SKU at every location, with an end result that is accurate at a global level.

Resilience in the face of disruption

Machine learning forecasting can be ‘always on’, to update automatically on the most recent data. New forecast accuracy and bias metrics can be calculated, the base and running forecast can be compared, and the updated results presented for review through dashboards. Forecast accuracy trends can be leveraged in adjusting demand planning. This ‘always on’ forecast monitoring, combined with dynamic and customer-level pricing and promotions, can be tuned to identify price sensitivity among customer segments, products that form a market basket, and thus build the foundations of an online recommender system. Once a daily forecast and customer history is merged with a transactional recommender system, the value of the recommender system in driving incremental purchases can be unlocked. This is where forecasting truly becomes an automated learning cycle. We have access to vast amounts of data and the challenge is to turn the mountains of information into actionable insights. Advances in forecasting algorithms means we can now sense and respond to changes in demand before the event becomes a risk to the business. An accurate view of demand helps ensure laboratories have the right supply of consumables to help deliver the products the market needs, when and where they are needed.

Author: Mark Balte is SVP of Product Innovation at Logility, logility.com