Improving Scheduling Efficiency at Zillow Offers
Co-Authored by Wei Xia, Applied Scientist, Zillow Group
Zillow’s new innovative home-buying program, Zillow Offers, provides a convenient and personalized home-selling experience for homeowners who want a certain and predictable sale on their timeline. We are constantly working to improve our customer’s selling experience, and one of the most important aspects of the home selling experience is efficiency.
To ensure the timely delivery of personalized offers to our clients, Zillow Offers relies on a seamless process that guides the home seller along the way. The process starts with Zillow Offers creating an initial offer based on the information of the home provided by the seller. Upon accepting the offer, the seller then moves on to schedule an evaluation with one of Zillow Offers’ professional estimators, who will evaluate the condition of the home and decide if any repairs are needed to complete the transaction.
As one of the more time consuming steps of the selling process, scheduling evaluations has become a bottleneck for the offer delivering phase, and coming up with a more efficient scheduling strategy has become a priority. Today, each seller is given the option to choose from three designated times on a future day, for example 8 AM, 11 AM, and 2 PM, for an evaluation. The three hour time slot is preallocated to allow for inspection as well as traveling between properties. To ensure more time can be spent on evaluations and less on the road, a mathematical model is devised to compute the best routes for estimators such that driving distance is minimized.
Our optimization does not stop with merely minimizing the driving distance. With the current scheduling strategies, the capacity of evaluations per day is around three times the number of estimators available. We want to also increase the evaluation capacity to better meet the growing demand for our service. We believe by increasing our scheduling capacity we can make our services more frequently available to more customers, which will also shorten the timeline of the selling process.
As part of an internship project, three new scheduling strategies are studied to increase daily capacity of evaluations, while limiting the potential impact on the customer’s flexibility. Both the improvement of total distance travelled and the evaluation capacity are considered, along with customer satisfaction for each of the strategies.
Exploring different scheduling strategies
Before reinventing the wheel, we conducted a quick survey on the current customer-facing scheduling services in the market. There are four main types of scheduling strategies, each based on how much control the service provider and the customers have on the scheduling date and time:
1) The customer selects a date and an exact time (we will refer to this strategy as “Zillow strategy”)
This strategy has the least uncertainty for the customers because they know the exact time the estimator will show up at their door.
2) The customer selects a date and a time window (we will refer to this strategy as “Amazon strategy”)
Remember the days when you needed to sit at home all day just to wait for the delivery man because he or she will come by sometime that day?
3) The service provider decides a date and the client chooses time window (we will refer to this strategy as “Wayfair strategy”)
Compared to the Amazon strategy, Wayfair gives the customer less freedom by providing a date, but still lets the customer decide when they want to be home to collect the delivery.
4) The Service provider decides the date and chooses a time window (we will refer to this strategy as “Best Buy strategy”)
The strategy with least freedom and has an uncertainty of arrival time.
For service providers, the efficiency of these services increases from strategy 1 to strategy 4, as the more uncertainty around arrival time, the more room they have for optimizing the schedule. However, Wayfair strategy and Best buy strategy largely restrict customer flexibility, while the Zillow and Amazon strategies do not leave enough room for optimization. We hope to choose strategies that increase the scheduling efficiency with a reasonable restriction on the clients’ flexibility.
As we have mentioned before, the improvement of scheduling efficiency sacrifices customer flexibility. Each scheduling strategy has a unique balance of scheduling efficiency versus customer flexibility, and the relationship can be roughly defined in Figure 1 below.
All strategies are trading off between scheduling efficiency and customer flexibility in some way. On one end of the spectrum, the service provider may dictate all the decisions and generate a very efficient schedule after observing all the requests, at the price of not giving customer any control. On the other end of the spectrum, the customers get to select the best time for them, but may force the service provider into a bad schedule which is not efficient in terms of both driving distance and number of customers served. We hope to choose strategies that increase the scheduling efficiency with a reasonable restriction of customer flexibility, so more customers’ homes can be evaluated sooner. One Natural question is, is there a strategy out there where both efficiency and customer flexibility can be improved from the current strategy?
Hybrid Strategies
Instead of adopting an existing scheduling strategy, we are trying to find a combination of some of the existing strategies to create a hybrid strategy. We hope to find a more balanced strategy that lets us serve more customers and at the same time maintain a good level of flexibility.
Hybrid Strategy 1: Customer chooses the date, Zillow decides the time slot
We still assume the three-hour-time-slot-per-day arrangement as the current strategy, but the customer only gets to select the date for the evaluation and Zillow will decide the time slot for them. The position of this strategy on the tradeoff plot is shown as the yellow dot in Figure 2 below.
Hybrid Strategy Two: Customer chooses morning/afternoon window, Zillow assigns a specific time
One of the assumptions that restricted scheduling capacity is the three hour time slot assumption. In practice, an evaluation time can take from half an hour to two and a half hours, depending on the home size and conditions. To explore opportunities around variable evaluation time, we consider the strategy where the length of each slot (evaluation duration + travelling time) is no longer three hours. Instead we allow different evaluation durations according to the house’s features. There are multiple ways to select the evaluation duration for each house, including random sampling of given duration length and prediction by machine learning model. Here we pick a random stratified sample from the historical distribution for evaluation durations, which are then randomly assigned to each client. Moreover, we do not preallocate one hour time for traveling between evaluations. Instead we use the actual traveling time between them, which is usually shorter than one hour.
Now with the variable evaluation durations, we can arrange potentially four slots per day, per estimator. We no longer restrict the number of slots for each estimator, and it is possible for them to evaluate four houses if some of the houses require a shorter evaluation duration.
Taking advantage of the variable evaluation durations, we propose hybrid strategy two, where the customers get to choose between a time window of either morning (8:00 AM-12:00 PM) or afternoon (1:00–5:00 PM), we then decide the actual arrival time of estimators for them. The position of this strategy is shown as the blue dot in Figure 3 below.
Hybrid Strategy Three: Customer chooses the day, Zillow assigns a specific time
Hybrid strategy three also assumes different evaluation durations for each house. There are potentially 4 slots per day, and the starting time of the slot is not fixed. Rather, the starting time of each of the 4 time slot is decided by the mathematical model. The customer only chooses the date, and Zillow will decide the arrival time of estimators for them. The position of this strategy is shown as the gold dot in Figure 4 below.
Building the mathematical model: TSP or VRP?
Talking about different scheduling strategies is easy. However, to measure the performance of each strategy, we need to use data and mathematical modeling to simulate how the strategies will work in practice. In other words, we need to create a virtual scenario where customers schedule inspections under each scheduling strategy, and see if some strategies allow us to come up with better routes for the estimators than others. Not surprisingly, finding “good” routes for estimators for a given set of inspections is a typical optimization problem. If you are thinking of the Travelling Salesman Problem (TSP), you are very close. The TSP is a classical combinatorial optimization problem that asks the question: Given a set of locations, what’s the shortest route that visits every location and returns to the starting point? Our problem is more general than TSP — it is in fact the well-known vehicle routing problem with time windows (VRPTW), which is a generalization of TSP.
The main difference is that the TSP aims to find the single shortest route, whereas in our case we need to find a route for each estimator so that the total distance traveled by all estimators is minimized. Additionally, the inspections have specific time windows for when the estimators should arrive (so that the homeowner is present), whereas TSP just requires the location be visited without any time constraint. There are sophisticated open source solvers that specialize in solving different variants of VRPs, and here we will use Google’s ORTools library* to solve our scheduling problem.
Result Analysis
In the test phase, the original model is revised according to the three hybrid strategies above. After testing each hybrid strategy with historical data, we summarize the results in table 1.
The performance of hybrid strategy one, two and three is shown as Figure 5 below. Here the x-axis represents different strategies, and the y-axis represents the average daily inspection slot improvements.
Conclusion
From the result analysis above, one can observe the number of throughputs is improved for all three strategies. Even if the current number of customers per day remains unchanged, the total traveling distance can decrease by as much as 38.0% in the new strategies.
These strategies will allow us to scale our business with excellence. We are able to schedule evaluations more efficiently to more sellers with less restrictions on seller flexibility. This means potentially shortened selling process for more of our sellers, and better utilization for our field teams. The future is bright, and with the power of mathematical models and machine learning algorithms, we believe we have found the key to providing better experience and efficient selling process for all Zillow Offers sellers
* OR-Tools 7.2. Laurent Perron and Vincent Furnon. https://developers.google.com/optimization/.
Many thanks to David Fagnan, Amy Jarvis, Sebastian Wickenberg, Bilgin Cemal for guidance and editorial help on this article. Thanks to the creators of the algorithm Suleyman Karabuk and Alex Pryiomka for providing valuable feedback on the development of the different scheduling strategies, and to the entire ZO Machine Learning team for their efforts getting us this far! Special shout out to Hannah Stombler-Levine and Megan Roewer for the graphics, and Mark Olson and Carol Anne Solberger for making this happen!
