« Back to Results

Machine Learning in Real Estate

Paper Session

Sunday, Jan. 5, 2020 8:00 AM - 10:00 AM (PDT)

Manchester Grand Hyatt, Nautical
Hosted By: American Real Estate and Urban Economics Association
  • Chair: Thies Lindenthal, University of Cambridge

Neighborhood housing rent index construction and spatial discontinuity: a machine learning approach

Yaopei Wang
,
Shanghai University of Finance and Economics
Yong Tu
,
National University of Singapore

Abstract

We integrate a hedonic housing rent model (econometric approach) into a state-space model (reinforcement machine learning approach). We adopt the kalman filter and smoother recursive algorithm and the expectation maximization algorithm (statistical estimation methods) to estimate the proposed state-space housing rent hedonic model. The method is applied to the Singapore public open rental housing market to construct housing rent indexes. Compared with the conventional econometric methods in index construction, the proposed model has three advantages. Firstly, a state-space modeling approach technically allows us to construct neighborhood level housing rent indexes through a reinforcement learning process regardless the sample size in a neighborhood. Secondly, the expectation maximization algorithm effectively enhances the robustness of maximum likelihood estimation for a dataset being repleted with unobservable information, for example, fewer or zero transactions in certain time periods. Thirdly Kalman filter and smoother recursive algorithm optimizes the estimates by capturing all information (before and after a time point) to predict a housing rent at a time point. This helps reduce the bias caused by sticky rents. The paper empirically proves that the proposed model outperforms other types of index models in prediction accuracy, hence produces more accurate housnig rent indexes at neighborhood level.

Accurately constructing neighborhood housing rent indexes are impotant in real estate valuation, real estate investment returns and risk analyses. This is because the spatial patterns of housing price distribution may change over time, which is resulted from urban developments. To illustrate it, we apply K-shape clustering algorithm in unsupervised machine learning literature to the neighborhood housing rent indexes to analyze the dynamic patterns of the spatial distribution of housing rents. We find the spatial discontinuity of housing rent dynamics. The housing rent indexes in some spatially disconnected neighborhoods appear to have similar dynamic pattern, while different dynamic patterns are found in some spatially adjacent neighbothoods.

Deep Learning for disentangling Liquidity-constrained and Strategic Default

Yildiray Yildirim
,
Baruch College
Arka Bandyopadhyay
,
Baruch College

Abstract

We implement a Deep Neural Network (DNN) methodology for disentangling liquidity-constrained and strategic default using a proprietary Trepp data set of commercial mortgages and motivate a model-agnostic interpretation of variable importance that is robust (insensitive) to severe Financial Crisis (2008). We use 29 loan-specific, property-specific, tranche-specific and deal-specific co-variates, 4 macro-economic variables, 2 indices for approximately 200,000 loans over a period of 19 years from 1998 to 2016 on a monthly basis. The non-linear activation function in the DNN captures the highly non-linear interaction among co-variates, geographical cross-correlation and macro-economic factors at state level beyond linear unobserved time-fixed and loan-fixed effect. We are able to capture strategic delinquency from the Variable Importance table and Shapley values for Deep Learning, e.g., Net Operating Income, Appraisal Reduction Amount, Prepayment Penalty Clause, Balloon Payment, etc. co-contribute and interact in a highly non-linear way to impact the endogenous choice of delinquency class compared to other more statistically significant variables such as Loan-to-Value. Further, we show that our results on variable importance are robust to the Financial Crisis of 2008. There is a significant increase in accuracy of predictions for the classes beyond 90 days of
delinquency when the Deep Learning Model is compared with Naive Bayes, Multinomial & Ordered Logistic Regression, Support Vector Machine, Distributed Random Forest, Gradient Boosting Machine, Deep Neural Network by gradually relaxing the specification structure, thereby increasing flexibility. These findings have significant implications for CMBS investors, Rating agencies, and Commercial Property Finance policymakers.

Contractual Completeness in the CMBS Market: Insights from Machine Learning

Brent Ambrose
,
Pennsylvania State University
Yiqiang Han
,
Clemson University
Sanket Korgaonkar
,
Pennsylvania State University
Lily Shen
,
Clemson University

Abstract

"A complete contract attempts to specify the rights and duties of the parties to the contract for every possible future contingency. While contracts are rarely, if ever, complete, there plausibly exists a spectrum of completeness. The degree of contractual completeness is particularly salient in the commercial mortgage backed securities (CMBS) market, where the Pooling and Servicing Agreement governs the actions of the agents who are party to the deal, and where investors have limited interaction with mortgage originators and servicers following deal inception.

Previous studies of financial contracting, which examine either the use of various contractual features or their welfare implications, tend to focus on a few specific features or paragraphs of a contract; thus excluding from the analysis information which is, admittedly, difficult to quantify.

We study whether commercial mortgage backed securities suffer from incomplete contracting problems by using new machine learning (ML) methodologies that are capable
of processing large quantities of textual data. Thus, we analyze how the completeness of Pooling and Servicing Agreements vary with features of the underlying the CMBS pools."

The Odd One Out: Asset Uniqueness and Price Precision

Carolin Schmidt
,
Centre for European Economic Research (ZEW)
Thies Lindenthal
,
University of Cambridge

Abstract

Based on applied machine learning (ML) techniques this paper suggests that round prices are not purely random events but are linked to liquidity and the uniqueness of the asset. First, using residential transaction data from the UK, we show that the availability of information from comparable sales influences the odds of observing a sale at a round price. Second, we explore ways to play to the strengths of deep neural network and incorporate computer vision approaches and building level imagery. Adding information on a building's vintage and the typology of its direct surroundings to the training data boosts the predictive power of the suggested ML classifiers. When a house is "the odd one out", its value will be relatively difficult to establish which implies that sales prices suffer from a relatively high signal-to-noise ratio. Automatic appraisal systems or index estimations could improve their accuracy by incorporating our findings.
Discussant(s)
Alexander van de Minne
,
Massachusetts Institute of Technology
Dominik Rehse
,
ZEW Mannheim
Tugba Gunes
,
University of Cambridge
Chris Cunningham
,
Federal Reserve Bank of Atlanta
JEL Classifications
  • C4 - Econometric and Statistical Methods: Special Topics
  • R0 - General