According to Loeiz Bourdic, Director Director at PriceHubble in France, real estate specialists like Gecina, BNP Real Estate, ING or even Axa Investments managers now want to get accurate estimates for the properties they are offering for sale with just a few clicks.
Continuation of the article below
To get there, big data and artificial intelligence are the partners of choice. PriceHubble works with BtoB clients (developers, investors, social landlords, etc.) and BtoBtoC (real estate agencies, finance banks, brokers, asset managers or even notaries).
In any case, this is the Swiss PriceHubble’s bet in 2016, the date it was founded. Since then, the startup has been represented in five countries (France, Switzerland, Germany, Austria and Japan), employs 60 people and has several hundred customers – including those mentioned by Loeiz Bourdic.
The Paris Chamber of Notaries wants a real estate valuation model
This calling card also convinced the Chambre des Notaires de Paris.
It evaluated four solutions in order to design an AVM (Automated Valuation Model) before choosing PriceHubble. “In my opinion, this is a relevant decision, as PriceHubble meets all the criteria that we set ourselves,” explains Stéphane Adler, Vice President of the Paris Chamber of Notaries and responsible for IT projects within this institution in Paris.
The Chamber of Notaries wants to use this algorithm for the Île-de-France notarized real estate data used by the Paris Notary Service. These data correspond to sales that have been recorded in Île-de-France for more than 25 years.
“We have digitized files that are sent to us by the notaries at the time of sale. This database collects information on goods and buyers: type of goods, floor, presence of a balcony, exhibition, type of buyer, etc. », Stéphane Adler list.
“The Paris notaries database is not perfect, but it is very well structured. It took us less than five days to use and integrate the first model, ”continues Loeiz Bourdic. “There is 25 years of data, but there is no need to integrate them all. Not all of them are relevant given the current market situation.”
However, this basis is not the only source of information for the design of this AVM.
“Today it is possible to add variables like transportation, schools, shops, and so on. By correlating this information with an algorithm, you will get a more accurate, efficient and faster estimate of the value of a property, ”assures the Vice-President of the Chambre des Notaires de Paris.
The final tool allows notaries to respond to customers seeking an opinion during a succession, donation, or an initial pre-sale meeting. “Your customers have an idea of the price of the property, but there are often surprises up and down,” says Stéphane Adler.
Such a tool also fulfills a need for development. “The Chambre des notaires de Paris had to deal with the free provision of the DVF databases (Demand for Land Value). The notaries wanted to keep adding value to their clients by going further, ”notes Loiez Bourdic.
Importance of explainability
“Notaries want a powerful and explainable algorithm. Loeiz BourdicPriceHubble
The notarial organization nevertheless had a very clear requirement. Transparency. Its members want to be able to understand the results of the property valuation model. “Notaries want an algorithm that is both powerful and explainable,” confirms Loeiz Bourdic. “But we are the opposite of a black box. This also convinced the Chamber of Notaries. “
PriceHubble’s French director believes that there are several ways to estimate the price of a property. The standard is the use of classic statistical methods. The adjustment of the variables is at the discretion of the statistician, who continues with his own analysis grid. Then it is possible to use generic algorithms.
“Algorithms like random forests and some neural networks are very powerful, but they are complete black boxes. You cannot explain their results. And if the output is inconsistent, you can’t fix it, ”he says. “Our approach is to offer something in between: We keep the explainability and control over the various variables. We are able to control the effects of several marginal variables. “
PriceHubble relies on “fairly advanced” set methods for this.
“The structure of the algorithm that we use in France, Germany and Japan is identical. The special features are integrated through the data that we enter into the machine, ”explains Loeiz Bourdic.
“In Japan there are special features related, for example, to the construction rules caused by seismic activity. The Chambre des notaires de Paris has its own confidential variables. It is the way of crossing data between them that is systematized in PriceHubble’s models, ”he adds.
The startup uses “elementary” stones like TensorFlow. “TensorFlow is much lower. This library makes it possible – subject to the initial investments, which are scalable – to maintain control over the algorithms, ”emphasizes Loeiz Bourdic. However, it does not rely on Scikit Learn, which does not allow the desired level of control over the algorithms to be achieved.
Behind the curtain, Pricehubble manages its ML pipelines with Spark, Docker, Kubernetes, and CircleCI. The model can then be integrated via an API, for example for a real estate agency who would like to benefit from an appraiser on their website. PriceHubble is also developing a user interface that makes it easier to access market research information. The user then manipulates a dozen different analyzes to understand market dynamics, building permits, prices, etc.
A “partnership” between notaries and data scientists
Regardless of whether it is notaries or PriceHubble, the two parties believe that this project goes beyond a simple supplier-customer relationship.
“We usually play the role of an external service provider offering estimation solutions. The project with the Chambre des Notaires de Paris is a real partnership. Loeiz BourdicPriceHubble
“We are used to working with institutions with which the relationship is one-sided. We usually play the role of an external service provider offering estimation solutions. However, the project with the Chambre des Notaires de Paris is a real partnership, ”says Loeiz Bourdic. “You [les notaires de la région parisienne] Provide your data so we can use a machine learning algorithm that is even more precise than what we are currently doing. “
“It’s a partnership and that’s how we really work,” adds Stéphane Adler. “At the moment this is a research and development phase. But if the solution works, there will be a continuation and maybe an association. We have the data, they have the technical capabilities. One without the other, it doesn’t work ”.
If the algorithm is effective, the solution is directly available to customers. “In all dense cities, we will usually get good results. This will not necessarily be the case in the countryside or in the holiday areas, “says Stéphane Adler, adding that in this case it will never be worth the human expertise that generates the visit to the property and the cost of repairing the property definitely. the apartment for example ”.
Second, the tool could also be used by certain financial institutions that lend to “real estate companies that want to get an opinion on the value of their assets every two to three years.”
Before the health crisis, PriceHubble and the Chambre des notaires de Paris had planned a four-month R&D phase. A job that started well in March 2020.
The aim of this phase is to exchange the different versions of the model between data science and real estate experts in order to improve its accuracy. “There will be different iterations of the algorithm that will be refined over time through the integration of variables and new approaches,” explains Loeiz Bourdic.
Despite the crisis, it is expected that “at the end of the year there will be something presentable,” says Stéphane Adler.
Other AI projects with notaries
The Paris Chamber of Notaries is not the first attempt at an AI. She actually has four projects underway.
Among other things, she is developing the VictorIA solution in collaboration with the Hyperlex company, the aim of which is to accelerate access to real estate assets and their classification in online data rooms.
In the future, the Chambre des notaires would also like to be able to search, recognize and, if possible, compare documents.
She had previously worked with Hyperlex to classify the use of real estate in Paris in a single database (VIDOC) in 1970. The aim is to determine whether the status of a property has changed (e.g. from a residential apartment to an office) in order to provide the customer with the history. The project is already three years old.