Walter Living’s data-driven approach to helping you make a winning offer
It’s the most common question our customers ask us: How much should I offer for a house? Some are afraid to miss out on their dream home, whereas others mainly want to avoid paying too much. It can be a difficult question to answer, especially if the aim is to provide objective advice based on data. Never one to back down from a challenge, Walter Living developed an algorithm to help us predict winning offers. This blog post explains our method step by step.
What’s the winning offer?
To systematically make winning offers, we have to be able to predict other people’s bidding behavior. If your bid is the highest, the house is yours. If your bid is lower than others, it will go to someone else.
We don’t have data on individual offers, but we do know what the winning offer was on each and every house sold in the Netherlands: its selling price. So, if we can predict the selling price of a property, we can predict the winning offer.
The question “How can we predict the selling prices of houses?” served as the beginning of our search for a data-driven answer to the question of how much people should offer for a house.
Selling prices of 380,000 houses
We took a dataset of more than 380,000 recently sold properties as the starting point for our research. We randomly selected several houses from this dataset and looked up the reference properties for each selected house. These reference properties are determined by our Automated Valuation Model (AVM), which looks at aspects such as the price per square meter, the asking price and other distinctive features of the selected house.
It turned out that the asking price is the strongest predictor of the selling price. There are two reasons why this is the case. Firstly, the asking price is often a realistic representation of the value of a property. Secondly, many buyers base their offer on the asking price. This means offers are informed by the initial asking price set by the real estate agent.
We found that the relationship between asking prices and selling prices follows a normal distribution. The image below shows this relationship in a neighborhood in Haarlem. As you can see, the pattern of selling prices closely follows that of the asking prices.
This pattern is the same for most neighborhoods, which is a positive thing. It defines the range within which offers on houses in a certain neighborhood tend to fall. And now that we know what this distribution looks like, we can begin to calculate winning offers.
There are currently various examples of neighborhoods in the Netherlands where the relationship between asking and selling prices appears to be nonexistent. There’s a clear trend of overbidding in most of Amsterdam, for example. But even neighborhoods where overbidding is common follow a predictable pattern, allowing us to tell you something about them, too.
Bidding strategies change over time
In addition to the normal distribution of winning offers in neighborhoods, our data showed another pattern: bidding strategies change over time. The graph below shows how much people overbid (in percentages) on a group of similar properties in Arnhem between 2015 and now. As you can see, overbidding increasingly occurs.
This upward trend can be seen in many neighborhoods and is probably caused by the current shortage of houses for sale in the Netherlands. As demand for houses has increased in recent years, buyers are making increasingly higher offers to get their hands on a house.
We use a linear regression technique to take the passage of time into account and estimate the effect of price trends of the preceding two years. This is how we make sure that old data points in neighborhoods where prices have risen don’t affect the offers we calculate.
Developing our bidding algorithm
Based on the above findings, we developed an algorithm that calculates three different bidding strategies using the following steps.
- Walter Living’s existing AVM collects reference properties in the area. For each house, it selects the 25 most similar properties from a database of recent transactions (currently containing over 380,000 transactions).
- As the housing market is constantly changing, we compensate for the “age” of transactions when determining the relationship between asking and selling prices. This prevents older transactions from affecting our estimate and making it too low in a neighborhood where prices are rising, or, conversely, making it too high in a neighborhood where buyers have begun to adopt more conservative bidding strategies.
- Based on these 25 transactions, we can tell you something about the normal distribution of the relationship between asking and selling prices in the neighborhood in question.
- We can now tell you where each offer falls on our estimated distribution and what your chances of success are with each offer.
The numbers tell the tale
Our method may sound like it makes sense, but we could only make sure it actually worked by testing it. As we at Walter think it’s important that our data and calculations are as accurate as possible, we extensively tested our algorithm.
We did this by checking reference properties against 10,000 houses from our dataset.
- The “test set” was only used to test the algorithm and was subsequently removed from the database for the duration of the test. The test would be successful if the algorithm produced the winning offer as often as it should for each strategy.
- For each test transaction, we had an offer based on the asking price and the algorithm. If the algorithm produced an offer that was higher than the selling price, it was a winning offer.
In this test, we checked whether a strategy that should yield a 50% chance of success did actually result in a winning offer 50% of the time. This means that, for this strategy, about 5,000 of the 10,000 test transactions had to produce a winning offer.
We want to give our customers several bidding strategies to choose from. That’s why we made the test even more difficult for our algorithm to pass. We offer our customers three strategies: “Play It Safe” (50% chance of success), “Play to Win” (85% chance of success) and “Go All In” (95% chance of success). These strategies provide insight into how much higher your bid has to be (in percentages) without it falling outside the normal distribution. In that case, you would have to pay more than buyers usually do in the neighborhood in question.
We tested each strategy separately, just to be sure. “Play It Safe” targets 50% of the houses with a conservative bidding strategy, “Play to Win” aims to produce a winning offer 85% of the time and “Go All In” aims for a winning offer 95% of the time. With 10,000 test transactions, this meant the algorithm was expected to produce a winning offer 5,000, 8,500 and 9,500 times, respectively.
It’s interesting to see that the amount of extra money offered isn’t equal to the increased chance of getting the house. Although the difference in the amount of money between the different strategies is the same, the increased chance in the first step results in an additional 25% chance of success, while the second step only results in an additional 10%.
Our bidding strategies work
The table below shows the results of the three strategies. Our predictions correspond with reality. The chances of making a winning offer on a house largely match the predictions. Our model is this precise because it only uses one variable: the relationship between selling prices and asking prices.
Walter Living offers you a tried-and-tested bidding strategy
The Walter Living algorithm gives our customers a clear idea of their chances of success with a specific offer. Our algorithm uses relatively little information to provide a good indication of the range within which a house will almost certainly be sold.
Deciding how much money to offer is ultimately up to the buyer. If you’ve found your dream home and you want to buy it at any cost, “Go All In” is the best option for you. If you’re not particularly attached to the house and you intend to move again within five years, “Play It Safe” is probably your best bet.
Of course, a house may fall outside the parameters of our algorithm. This especially goes for atypical houses with few good reference properties. That’s why we always have a Walter expert stand at the ready when you put together your offer. Our experts can find all relevant information about the house within seconds and give you the human interpretation of the data.
