The Challenge of Market Selection - and Cadre’s Approach

“Location, location, location.” The value of real estate is in large part a function of location, and forming an investment thesis regarding which markets to target is paramount to real estate investment success.

Given the importance of market selection to investment performance, it’s no surprise that we spend a lot of time thinking about how we can leverage data to gain an edge on this front. Historically, the opaque nature of commercial real estate markets has meant that such subjects rarely get treated quantitatively. However, we believe that the increasing availability of traditional and alternative data sources, combined with recent advances in machine learning, allow for a more transparent and thoughtful approach to market selection.

To understand the impact of market selection on investment performance, consider the following anecdote based on real-world property transactions.

Here we see two multifamily buildings that, at a high level, share many of the same characteristics, save for the market they’re located in. Each building is sold twice during the same two time periods. Property A in Houston only appreciates by about 7% over the roughly five-year hold period, while Property B in Atlanta appreciates by close to 29%. With the benefit of hindsight, we can find many such examples where market-level dynamics drove differentiated returns between otherwise similar properties. This of course begs the question: how can investment professionals anticipate and profit from these sorts of market dislocations?

The Conventional Approach

Existing approaches to market selection in the commercial real estate (CRE) industry typically fall into one of two categories:

1. Expertise-driven selection strategies: Smaller firms will often pursue individual deals in an ad hoc fashion, based on relationships and prior experience. Some form of market analysis may be done when underwriting the investment, but deal sourcing is mostly a product of intuition, past experience, and existing sponsor and broker networks. While the role of human intelligence and expertise shouldn’t be diminished, we believe that this informal approach to real estate investing is bound to lose ground to more systematic methods, just as it has for other financial asset classes.
2. Heuristics-driven selection strategies: More research-oriented shops will commonly rank and screen markets according to metrics like population or job-growth, with the implicit assumption being that these factors are the most important. These kinds of rules-of-thumb sound plausible, but they are rarely able to explain - much less predict - market performance. For instance, Houston actually had a higher average rate of population growth than Atlanta for the five-year period preceding the initial 2009 transaction in our example.

At Cadre, we believe that these approaches to market selection are insufficient. Consistently identifying the kind of market-level opportunities that lead to outperformance requires a systematic approach - one that’s powered by rich datasets, robust models, and the ability to rigorously test many different hypotheses on what factors drive market performance.

Cadre’s Approach

Cadre began by asking a simple question: “What causes certain markets to outperform?” Forecasting market and property-type price appreciation stood out as the most direct way to solve for this. CRE price levels have strong, causal relationships with demographic and economic variables. Additionally, the high barriers to transacting in CRE cause information to be incorporated more slowly into the marketplace. Whenever prices are both predictable and inefficient, there’s an opportunity for savvy investors to profit.

At a high level, our approach uses many, many instances of repeat sale transactions - like the ones shown in the example above - to identify the market characteristics that frequently presage higher relative returns. The relationships identified are then tested to ensure that they generalize across time in statistically robust ways. This gives us confidence that these relationships capture real CRE pricing dynamics, and that applying them to contemporaneous market data will yield informed projections on future growth.

Treating market selection as a price series forecasting problem requires three capabilities:

1. A means of measuring aggregate CRE values by market and asset class, so that pricing dynamics can be observed.
2. A way to describe markets in terms of structured data that captures likely price drivers.
3. A modeling and testing framework that can functionally relate how changes in market data influence changes in market prices.

While we’ll provide an overview of each of these three capabilities today, in-depth treatment will have to be deferred to future articles.

1. Measuring CRE Market Values

Using a single measurement to track the value of a group of assets is usually referred to as indexing, and price indices are ubiquitous across finance. When the assets are similar and trade frequently, index creation can be as straightforward as just taking an average value of underlying assets. Unfortunately, CRE transactions are both relatively infrequent and nonhomogeneous in nature. This greatly complicates index construction and, as a result, the industry has traditionally sought to track market values by aggregating appraised or self-reported asset valuations. These techniques often paint an overly smooth and lagged representation of true market prices.

‍
At Cadre, we believe that the best way to measure changes in CRE values is to look at the prices at which buyers and sellers actually transact. To this end, we’ve developed a suite of transaction-based indices using a proprietary repeat-sales factor model. Not only do the indices produced increase our coverage in terms of the number of geographic areas and asset classes we’re able to track, but in out-of-sample testing they explain a high-degree of the total price variation in repeat property sales. This ability to more accurately track a wider variety of price series in the CRE marketplace underpins our ability to model market appreciation.

2. Market Profiling

Where index creation is about representing many assets with a single series, representing a market is about taking a single abstract concept and developing a concrete description for it in terms of many different data series. The more the data, and the more relevant this data is, the more powerful this market description becomes. It’s for this reason that we’ve spent years building systems to ingest and harmonize data sets from various data providers. Our market modeling today incorporates over 40,000 variables and more than 3 million data points, and we’re working hard at adding new sources everyday.

3. Modeling and Testing

Time series forecasting is a well-studied subfield of statistics, and there are many different ways one could seek to model price appreciation. In this particular setting, we have a large number of data series and there’s an economic basis to expect casual, temporal, and spatial associations to exist between them. Whether and how to exploit such associations when forecasting prices is a fundamental design question. One could choose to ignore them and focus on univariate price models, use intuition to handpick a number of variables in econometrics fashion, or open up the problem to machine learning.

Cadre’s decision to use machine learning is motivated by the field’s recent success in time series competitions such as the 2018 M4 competition. Specifically, we use a form of deep learning based around recurrent neural networks (RNNs) in order to produce forecasts over a two-year horizon. RNN architectures are flexible in that they admit multiple sequences of data as input, and powerful in that they can generically model data interaction within and between these series. In practical terms, this means that our forecasts benefit from cross-series learning among a wide variety of likely price drivers.

The Path Forward

It’s important to acknowledge that any market selection methodology will have its flaws. Real estate data can be inherently messy and slow moving. What’s more, the pandemic has dramatically slowed transaction volume and put significant stress on a number of asset classes and markets - making the task of forecasting increasingly difficult.

Still, as we’ve discussed here, the strength of a machine learning solution is its potential to learn from the past to predict the future. How does our model behave when we’re in a once-in-a-lifetime event? We don’t yet know. It will be some time until we understand the true implications of the pandemic on our economy. That said, leveraging data becomes even more important when navigating uncertain times. Systems that monitor employment trends, rent collections, and transaction activity, combined with programmatic insights from our portfolio, our experience and our network of operators, should help.

Selection of the right market lies at the heart of any successful commercial real estate investment strategy. To outperform in today's environment, an investor needs to have a market selection methodology that isn’t weighed down by human biases, isn’t limited by inadequate data, and instead is robust enough to assess any location’s current state and future trajectory.

Cadre’s market selection approach - machine learning paired with on-the-ground investment experience - is rigorously designed to identify the high-growth markets that are expected to outperform. The outcome of this approach has been the development of The Cadre 15, a selection of top national markets we believe are likely to outperform in the current environment. The Cadre 15 includes Seattle, LA, Las Vegas, Phoenix, Denver, Dallas, Austin, Houston, Nashville, Atlanta, Tampa, Orlando, Miami and South Florida, Charlotte, and Washington DC.

To continue outperforming, it’s important to understand that market selection is an ongoing process - quality opportunities outside of focus markets can and do exist. A rigorous and scientific selection process is what will identify new opportunities as they arise.

Peter is a Staff Data Scientist at Cadre. He was previously at JP Morgan, working as a quant on the electronic credit trading desk and data scientist for the Debt Capital Markets group.