Dynamic Pricing for Dynamic Times

Many of us rely on ride-hailing apps (Uber, Grab etc.) for our daily transport needs. A unique feature of these apps is the rapidly changing price for the same ride, depending upon variety of factors including supply-demand. As a result, it is not uncommon for riders to try the same ride a few times to get a better price.

A central component of today’s on-demand economy is dynamic pricing. As we see more industries disrupted, we should see greater focus on getting this central component right. A Harvard Business Review (HBR) article estimates that on-demand economy consists of 22.4m consumers with USD ~58b spending annually. That is huge!

Revenue management via dynamic pricing is an established practice followed by airlines and hotels. It is also finding roots in retail/eCommerce and, in recent times, in ground transportation.

Ride hailing apps (Uber, Lyft, Grab etc.) utilize dynamic pricing heavily and regulators (example: Singapore’s Land Transport Authority) are giving taxi companies, which operated with fixed pricing, a free-hand in varying their fares based on supply-demand factors.

As consumers, we love a good bargain and dynamic pricing gives us those deals. The flip side, however, is that we may sometimes end up paying more than what we think is ‘fair.’

Ride hailing apps, in particular, have been in the news in various countries for excessively high surge pricing. They are trying hard to fix this by ensuring that the pricing models work well.

Types of dynamic pricing

A recent study done at University of California, Berkeley focuses on building automated real-time pricing engines that work well in on-demand economies. The key challenge in creating these models is rapidly fluctuating supply and demand, often from minute to minute. The objective of pricing is not only to maximize firm profitability but also boost the supplier and customer loyalty.

There are multiple pricing schemes, from simple to complex, with varying degrees of effectiveness in terms of profitability, retention of suppliers and customers.

1. Fixed pricing: Like a traditional taxi. This fixed price does not consider any supply or demand factors at all. This is a utility-like way of thinking about transport – the old way.
2. Proportional pricing: Price is equal to the base price multiplied by a constant factor, times the ratio between demand and supply. So, when demand outstrips supply, the price will go up. This is similar to surge pricing that we observe with ride hailing apps. The principle here is that when price goes up, more supply enters the market. Also, the high price moderates the demand resulting in a new equilibrium.
3. Probabilistic pricing: This is an approach based on real-time experiments, where the model starts with a base-price and uses a normal distribution to offer a few prices centered around the base. Then the base price is moved towards the price that was accepted by buyers using a learning rate.
4. Reinforcement learning: Reinforcement learning is like human learning. The autonomous pricing agent decides the next action based on an objective of maximizing a goal – in this case profit, supplier and customer loyalty. There is a discount rate that determines how much weight to put on immediate gratification vs. long-term objective. Based on the setup, the reinforcement agent can decipher strategies that work and applies them to the present situation.

So, which approach is better?

A simulation done, as part of the Berkeley study, demonstrates the performance of these pricing mechanisms.

1. Reinforcement learning based approach excels in short-term and long-term simulations on KPIs of firm profit, earning per driver and does a good job of matching commuters to rides.   
2. Surprisingly, the proportional approach, most likely in use today by major ride hailing apps, does only marginally better on KPIs than just having a fixed price.
3. Probabilistic model yields better firm profits but results in approximately the same driver earnings.

This study suggests that there is potential for ride hailing apps to explore Reinforcement learning based autonomous pricing agents, which are essentially a form of Artificial Intelligence, to aid their KPIs.

Infrastructure considerations

Reinforcement learning agents, considered here, utilize Markov Decision Process (MDP) which is computationally heavy. This is where GPUs outshine their general-purpose counterparts, CPUs. In other simulations, utilizing MDPs and GPUs, a speed gain of 90x is observed over CPUs alone.

As the on-demand economy replaces traditional industries, and as such Reinforcement learning agents become the real-time pricing experts that everyone relies on to provide the most efficient price for all parties involved, we will observe greater reliance on specialized infra with GPUs.

IBM’s PowerAI platform offers tremendous benefits in such situations and has potential to be the backbone of the on-demand world.

PowerAI boasts of 4 Nvidia Tesla P100 GPUs along with NVLink interconnects, Power8 CPUs and a very high memory bandwidth. This is coupled with open-source stack running on Ubuntu consisting of most popular deep-learning frameworks (such as TensorFlow, Caffe, Torch, Theano etc) and libraries that are optimized and precompiled. This offers a complete solution for the data scientists to run hundreds of experiments and optimize the models frequently.

Additional details on PowerAI can be found here:

https://bit.ly/poweraiblog