Market pricing: the modelling challenge

How can you accurately estimate real market prices when many transactions do not reflect a property’s true market value? Adrien Bernhardt, CTO of Homiwoo, looks at the various challenges involved in dealing with this vitally important issue and the solutions being deployed.

To build and then further develop its models, Homiwoo uses multiple sources of data among which the property sales recorded on a French government database (the DVF). The sales information is published twice a year with a four-month delay and covers every property transaction in the country, including housing developments and land, whether built on or not. 

The market price: where supply and demand meet

However, this search for information is just the tip of the iceberg. “At Homiwoo, what really interests us, and represents our biggest challenge, is not to provide the price indicated by the DVF database but to give people the market price,” explains Adrien Bernhardt. “In other words, it’s about accurately working out the price that balances supply and demand.”

While the prices of the vast majority of properties for sale are made public, sold prices are not necessarily a true reflection of the market price. The time taken to complete the sale needs to be factored in; properties that take a long time to sell usually achieve prices above the market value, while quick sales can be conducted at below market values. Demand also raises other sorts of challenges – starting with the fact that there is no database available to measure demand. On top of that, some transactions in the DVF database do not reflect the true market price of a property for a variety of reasons: in cases such as the disposal of joint assets after a divorce, an inheritance sale or vendors facing financial difficulties.

“In our job, the challenge is to integrate all these different aspects and then decide the true market value of a given location, so that the client can decide how to market their property,” says Adrien Bernhardt. “We have to allow for the fact that none of the data sources available gives a very precise figure on the true market value.” As a data scientist, he believes that more than a third of all properties on the market have significant discrepancies between their advertised and actual market values. To give one example, the sale of a flat owned by a social housing association to one of its existing tenants will involve a discount on the market price. So, if a social landlord asks for the market price of an address where tenant sales have been made, there is little point in providing the actual prices achieved – as they are not representative of the market.

An unrivalled market approach

While other market players use a dataset which they optimise by removing errors, Homiwoo has opted to develop a unique scientific approach that harnesses the potential of Artificial Intelligence and Big Data. “We have developed an unrivalled approach, specifically designed to meet the needs of property market businesses,” says Adrien Bernhardt. “Our model strikes the right balance from the various data sources available to reflect the real state of the open market and enables clients to market a property in an optimal way - with its price fully reflecting the laws of supply and demand.” The tools deployed by Homiwoo offer access to a wide range of indicators (price, time to sale/completion, type of market, etc.), granularity and contexts, which together provide clients with an extremely precise price range. The data is subsequently handled in a way that enables valuations to be adapted to specific use cases. To meet the varied requirements of social housing landlords, for example, it can provide the sale prices of previously rented properties, current rents on the open market, the price of new-build social housing, etc.

It's hard for Adrien Bernhardt to illustrate everything that goes into achieving this objective. “Although most of the Machine Learning algorithms involve supervised learning, the fact that there are no datasets for the reality we need to model means that we have to carry out some highly complex calculations. By comparison, using Machine Learning to work out the future pricing of a property between the moment it is brought to market and a given time doesn’t pose a particular problem. Even during such a volatile period as the current one, the issue of price variation is much less complicated to model.”

To build an accurate model, several approaches are used. One particular task is to identify errors in the source data and correct them, such as surface area figures that are miscalculated, features or addresses that don’t match, etc. Other data can be largely irrelevant or unusable when calculating the market price of a property. As Adrien Bernhardt sums up: “Homiwoo’s value-added is to draw on its market expertise to understand the true situation, and then model that … using data that does not reflect that reality.”