Wheelhouse’s Predictive Demand Model (Pt 3)

March 30, 2021 | Andreas Buschermohle

This post is part 3 of our Pricing Engine Overview. If you would like to read that first, you can do so here. Or, if you would like to read about Wheelhouse’s Base Price model first, we are sharing that here.

In the pages above, we detailed how Wheelhouse determines the Base Price for each unit, and also how we more deeply analyze and understand both markets and subsets of units.

Now, let’s dive more into how we isolate drivers of local demand, and how we leverage those signals to create far-future price recommendations before booking patterns emerge (e.g. how should we price a unit 365 days from now, when only <1% of a market has booked?).

The first local demand model we employ is our ‘predictive’ model, which is intended to predict accurate prices before a market begins to book. While this model has a number of goals, its most important job is to ensure that we do not underprice units for far future stay dates.

This model is particularly important for the role it plays in pricing local events and holidays that start booking many months (or years!) ahead of a stay date, and offer the best times to capture revenue.

For our Predictive Demand Model, we analyze prices in markets, including future and historical pricing data from hotels and short-term rentals. These pricing signals contain local and historical insights that we can leverage to better ensure the far-future is accurately priced.

Additionally, to aid interpretability and effectiveness, our model can parse these pricing signals, to determine the ‘why’ behind future price increases, so we can better understand factors, including:

Let us explore each of these factors in a bit more depth.

Determining Seasonality

How we extract seasonal patterns from market pricing data

When analyzing a market, our first goal is to accurately determine the ‘seasonality’ of a market. In most markets, this seasonality curve changes slowly from day to day, and has traditionally been pretty stable from year-to-year.

Let us examine a chart that shows the average prices in San Francisco by day of year. In this chart, the dots represent the average price by day of the year. The line shows the output of our custom-designed low-pass filter that extracts the seasonality curve.

Via this method, it is clear that the high season in San Francisco is from July through October, while the rest of the year is relatively flat, in terms of seasonal demand.

Note: Pre-Covid, essentially all markets had a unique, but identifiable seasonality curve, which has been quite stable across multiple years. Post-Covid, it is too early to tell if and when traditional seasonal travel patterns, in the US and other markets, will return.

Day of Week

How weekday prices differ by market

With a seasonality curve in place, we next want to analyze pricing patterns to discern a regular, repeating pattern around ‘day of week’ pricing. This analysis is particularly important in urban markets that have meaningful business traveler demand, and for leisure markets where we often see bookings for Fridays and Saturdays increase.

To better understand this, let us examine a ‘weekend effect’ as shown in the following bar chart of ‘day of the week’ pricing for San Francisco and Austin. To better compare the two markets, we show the daily prices relative to the lowest price and how much more other days increase above that minimum.

As you can see, San Francisco’s isolated day of week pricing is mostly flat with a very modest weekend increase of 3%. However, Austin clearly shows regular high demand for weekends, with most units increasing prices more than 50% on an average weekend.

By knowing this pattern for each market, we can again more accurately interpret future pricing patterns, and accurately name these factors for our customers.

Local Events

How we localize the impact of events

Lastly, our Predictive Demand Model leverages prices to detect local events, in the far future.

While similar to our other two demand ‘filters’, this model additionally considers the location of each event when understanding how particular price signals should be translated to the broader market.

Unsurprisingly, most events only impact pricing for units proximal to the event. However, in most markets there are multiple annual events (e.g. a huge conference, event, or holiday) that are so large they drive price changes across an entire market. Due to this, our model needs to be able to both (a) identify hyper-localized pricing signals and (b) translate those signals into an estimate of the scale and breadth of each event.

We achieve this by analyzing local pricing signals, removing the Seasonality and Day of Week impact, before extracting ‘short’ price increases. These results are achieved via a custom designed high-pass filter. The result of this filter is a highly readable chart that allows both our team (and you!) to examine any market for event-driven demand.

As an example, let us examine what our model determined to be the ‘event impact’ for two different areas in San Francisco, Union Square (near downtown) and Golden Gate Park (4 miles from downtown, a popular concert area). (Note that these are projections only from our Predictive Demand Model, as of January 2020. More on this soon!)

In the first chart (above), we see event-driven demand in the Union Square neighborhood, which is very close to the Moscone Convention Center. Therefore, we see our Predictive Model reacting very dynamically over the year, and ultimately dwarfing the relatively small seasonality impact in this area.

However, in the second chart (below), we examine units near Golden Gate Park, 4 miles from downtown. In this chart, we can clearly see that the biggest events were projected to be the two music festivals — Outside Lands and Hardly Strictly Bluegrass. These events do not impact the downtown area. And, only Dreamforce, the largest of conferences, can be definitively tied to a specific price spike in the Golden Gate Park area.

Importantly, our model is also able to leverage this signal to predict when significant demand is going to spill over into other neighborhoods, albeit at a smaller price multiplier.

Previous Post Next Post

TAGS: