Home
RM Tips and Strategy
RM Tips and Strategy

AI demand forecasting requires human learning

How to increase the results with a better technology?

Emmanuel Scuto
June 7, 2021

In front of a sophisticated technology, we can respond in two different ways. The first one is to directly dismiss the product, stating, “How can a machine do better than me”, without even testing it. The second one is to blindly use it without any critical thinking, accepting all recommendations by default, even important ones. In both these situations, the outcome cannot be satisfactory for humans.

Working in car rental yield management for decades now, I have often found these attitudes facing the technology developed by our Product & Innovation Engineers. When we happen to present the product to new clients, instantly, some of the audience is enthusiastic and is willing to use it without challenging the features. The other part of the group listens politely to the speech until it ends and leaves without giving it a try or asking questions. Then, the Customer Success Managers need to assist each of them to guide them through their experience with the product, and depending on their reaction to software, they act differently.

This phenomenon is a strange but common reaction of the brain. When faced with something new, the brain instinctively senses danger and refuses to accept the change. But often, our users do not even know they are supplied by an AI (artificial intelligence) using advanced machine learning features.

How to increase the results with a better technology?

In 2014, WeYield developed a technology to aggregate the results of market shop price collections supplied by IT experts like QL2 or Rate Gain. However, every single website has its own way to display the car descriptions, which need to be read and grouped. So, to match the users’ fleet requirement, we had to clean the description so that the system would suggest a common car description to be read and analyzed. With more than 250,000 different car descriptions collected, there is no doubt a human cannot manage them all. Sometimes we aggregated more than 200 different descriptions for a Toyota Yaris 2 doors with air conditioning depending on the source of the data scrapped! Imagine how many combinations it makes per model, per brand, per country, per website… Every new car description would have to be populated by every user and they would have to manually correct it and assign it to the proper group.

This makes me remember a lady I knew: her job was specifically to collect rates from different websites and clean them all manually. Each day, it took her from 3 to 5 hours to complete this task. It is completely unacceptable for us because of the WeYield philosophy: we want to give more freedom to our users and save their time!

AI demand forecast will not replace the yield manager but will work for him/her

In 2018, we set up an AI-based on machine learning to automatically read car descriptions, compare them with an existing reference database (that is continuously enriched from the inputs recorded every day). Because guess what? An AI can only learn based on existing references. There is no magic but only efficiency and automation. With the assistance of our AI, the car description is displayed almost, and it saves so much time for our users! The only remaining task is to group the cars they want into their own fleet group. That is specific to each of them.

Artificial Intelligence has a major benefit for WeYield yield managers. But it is also extremely important that the user interacts with the machine to give it some “instructions” to learn better. They can act as a teacher for the machine.

In 2019, the WeYield Product & Innovation team launched a new yield management project to have a demand forecast built thanks to AI technologies. We hired a team of data scientists and they all worked with Christian Cadéré who has a Ph.D. in Computer Science. After months of hard work and hundreds of different tests on various models, we selected the open-source algorithm called Prophet (developed by Facebook), because it gives overall the best results in the context of the car rental demand history, which has recurrent patterns.

So, we have started using it in production, testing it with beta testers within our WeYield client community. The first results have been extremely encouraging, and thanks to user feedback we managed to improve even more the forecast of demand, typically for a 30 days-booking window, which is the most important period to act on. In the yield management for car rental, usually 50% and sometimes more of the demand is confirmed during this range of reservation dates.

Even the best forecast must be challenged

However, a typical effect of an innovation providing better and faster results can also generate some negative and irrational impacts on a user’s brain. “If a system can do it better than me, then what is my role? What is the use of my job? What is my contribution to the company?” This is where the interaction between the machine and the human starts to give its maximum results.

When the conditions of the market change, sometimes too quickly, a human is the only one being able to analyze the new environment and adjust his decisions. For the system, no matter the quality of the algorithm, it will take days or weeks to “digest” the new trend and adapt its response. Christian Cadéré wrote a post to illustrate this problem in April 2020, during the first Covid-19 lockdown in France.

Facing this incredible and exceptional interruption of the reservations, we had to manually adapt our algorithm to the new uncertainty of the environment, because no algorithm can “understand” such a seismic change on its own. The system can only be as good as the human who challenges the recommendations and criticizes them for its permanent improvement. Nothing works perfectly on the first try. And it will remain the same if the user has the persistence to realize these action controls and doing so, increasingly improve the system.

An automated demand forecast in the yield management for car rental is not an end. It is a step to use its output and prepare the automatic recommendations in prices and fleet control. But even the best algorithm will have to operate between a set of rules, set up according to the revenue manager’s experience and knowledge of his market, to avoid crazy actions.

In the end, to get the most out of an automated system, don’t use it as a Blackbox, without any understanding nor control from the business analyst or the yield manager. Because they must TRUST it. And to build this trust, you must go into it, test it, take actions, measure the results, come back to Christian’s team, and discuss the corrected actions and help the system evolve, grow, and adapted to your needs. There is no magic but work with tech and human cooperation.

Published by
Emmanuel Scuto
Linkedin logo

Expert in Revenue Management and Pricing in the Car Rental industry for 20 years, I aim to share my optimization experience with our customers throughout the world. I am specialized in revenue maximization, pricing strategy, yield management, reporting based on AI.

Our readers also liked

MAIL LIST

Newsletter

Subscribe to our newsletter and be the first to receive the most relevant information and best practices on revenue management.

Thank you! Your submission has been received!
Oops! Something went wrong while submitting the form.