AI deep-learning model predicts optimal prices in uncertain times

But these predictions become unreliable when circumstances differ from the time the source data was generated, such as when the Covid-19 pandemic disrupted supply chains and altered consumer demand. 

In a study published in the academic journal Artificial Intelligence for Business, University of California Riverside School of Business professors Max Joo and Hai Che, and their collaborators from Baruch College and Ohio State University, also in the US, believe they have solved this problem by developing a model that considers both historical sales data and the economic theory of demand. 

The model is detailed in a paper titled ‘Theory-Regularised Deep Learning for Demand-Curve Estimation and Prediction’. 

Such a combination allows AI to quantify the unpredictable aspects of how prices affect consumer behaviour during unprecedented circumstances, explains Joo, an Associate Professor of Marketing. 

For example, demand for hotel rooms is typically higher during the peak summer months despite higher rates. A standard AI model may mis-predict that higher prices are associated only with higher demand. 

However, affordability limitations and sense of price fairness will still limit how much holidaymakers are willing to pay. Such factors are difficult to gauge during uncertain times, Joo says. 

The new model aims to bridge a gap between standard AI models and the real-world complexities that businesses face when setting prices – particularly in unprecedented scenarios such as economic shocks and extreme price fluctuations. 

To validate their model, the researchers analysed pre- and post-Covid retail data for breakfast cereals, which had a sales surge during the beginning of the pandemic, but later returned to a pattern of historic sales declines.  

They compared their new model, combining economic theory with standard deep-learning and log-linear models, and evaluated their ability to predict demand changes as prices fluctuated beyond historical ranges. 

Study produces compelling result 

The results were compelling, the researchers say. While standard models performed well with data within the known range, their predictions deviated when confronted with post-pandemic price levels. 

The researchers’ model, however, retained high accuracy, demonstrating a substantial improvement over other methods by reducing generalisation errors by up to 50% in some cases. Such errors occur when a model trained on a specific dataset does not perfectly capture the underlying patterns or relationships that are held in different contexts. 

“The pandemic was a perfect stress test for our model,” Joo states. “The price and demand patterns during Covid-19 differed significantly from any prior period. This was exactly the type of scenario where typical AI models would struggle to produce accurate forecasts.” 

While most AI models that rely heavily on past price data falter when circumstances change, the new model’s ability to utilise economic theory gave it an edge, Joo emphasises. 

“We’re combining the best of both worlds – advanced AI techniques and established economic principles – to create a system that’s both intelligent and adaptable. The future of AI in business isn’t just about exploiting more data; it’s also about leveraging human knowledge from various domains to build smarter, more reliable tools.” 

You can find out more about the study here.