The complex dynamics of pricing

The definition of pricing dynamics is simple- the price changes per the intensity of the real-time demand for the product. The price change of any dynamically priced product is derived depending on the time of sale, product inventory on hand, and numerous other factors. Hence, leveraging analytics to execute pricing dynamics optimally is not that simple.

Like many other methods, we have tried to bring more and more science to pricing dynamics. While bringing quantitative approaches helps to some degree, after a certain threshold, these approaches start hurting. The reason is simple. With the current approach that we use for pricing, we do not capture some of the “softer” aspects effectively.

Pricing dynamics is very strongly impacted by factors like (examples, not exhaustive list):

• Dynamics within the category
• Consumer confidence
• Economic forces (macro and micro)
• Consumer behavior
• Actions of market competitors
• Brand positioning

To add to the complexity, most of the variables listed above operate in different time scales, are subject to different rules, and are not great candidates for extrapolation or mathematical modeling. To build a software solution that can effectively work for pricing dynamics, it needs to have the following characteristics (again, not an exhaustive list):

• A wide basket of scenarios and associated models
• Pricing control algorithms
• Real-time volume estimators
• discount heuristics (all time periods, including real-time)
• Consumer behavior model
• Market dynamics

Can Deep learning help?

I strongly believe it can. In fact, I believe that is the only way.

Due due to the wide gamut of complexities mentioned above, I believe majority of solutions today, though they try to bring as much science into this as possible, still leave a wide-gap.

The traditional solutions available today have their roots in decades-old pricing methodologies, from a time when practical applications of neural networks were far away from feasible. Most of these solutions could not capture all the factors because it was impossible. The problem is they still do not capture it all in an age where it is possible to incorporate most of these factors precisely. That is also the reason most leading retailers build their own pricing algorithms.

Building a dynamic pricing solution leveraging deep learning is very much possible now. It will not be easy and is not for everyone. The combination of people, processes, and technology needed, at scale, to execute something like this successfully means that only very large companies can build something like this (at least during this decade).

The most significant legwork for building this type of model is actually:

• Capturing all input factors, examples of which were mentioned above
• Developing child-models to quantify these factors, the inputs of which will feed into the parent model (training data at first)

If you can build these successfully, you are destined to successfully build the deep learning model for pricing dynamics optimization. The reason I say this should be obvious. Think about transforming brand positioning, competitive actions, et al. into models. System dynamics methodologies have been tried to capture and quantify these, but there are better ways to model them. But once you can model these, you have already won the battle.

References:

• Pricing, Tudoe Boeda, Business Expert Press
• Pricing with confidence, Holden and Mukherjee, Wiley
• The Strategy and Tactics of Pricing, Angle et al., Routledge