Putting Price Tags on Houses Survival Analysis with Real Estate Data Marco Salvi Financial Engineering Real Estate marco.salvi@zkb.ch
2
ZKB and homegate.ch: Who we are! 3rd largest Swiss bank 70.9bn CHF total assets 110bn CHF assets under mgmt! Strong in mortgage financing! Switzerland's largest online listing service >50% market share >17m pageviews per month! 100% subsidiary of ZKB 3
Overview! Real Estate Market: not simply bricks and mortar...! Emergence of data-driven valuation tools! Online listing services! New opportunities 4
Real Estate Market: Not simply bricks and mortar! Until the 90's: Virtually no data available Analysis dominated by 'gut feels' Listing exclusively through newspapers! Since then... New rich data sources / GIS Large-scale use of Automated Valuation Models (AVM) Online listing of houses for rent or sale 5
Automated Valuation Models: A new hedonist era! Purpose: to estimate the market value of properties by statistical methods! Technique: so-called hedonic modelling, known to academic economists since the 60s! AVM in a nutshell Collect transaction prices of houses and record their characteristics (size, age, location, quality etc.) Use regression techniques to extract the market valuation of the characteristics 6
AVM in action: A standard has been set! AVM first introduced in Switzerland by ZKB in 1997! Cost effectiveness: Traditional appraisal: ~ CHF 900 (EUR 600) Cost of an hedonic appraisal (all-in): ~ CHF 30! Precision: 60% of appraisals ±10% of market price! 2002: Approx. 40% of all Swiss mortgage applications evaluated with hedonic models 7
Online Listing Services: A success story! % of respondents using the Web when searching for a new home in Switzerland 50 45 40 35 30 25 20 15 10 5 0 1998 2000 2002 source: NZZ 8
Online Listing Services: Putting data to work! New opportunities for creating valuable (i.e. marketable) statistical feedback to Sellers / landlords: What is the price range that I can expect for my property / apartment? Buyers / tenants / listing service: What are the most interesting offers? General public / media: Real-time real estate indexes " www.homegate.ch/index starting August 2003! Opportunities depend on the peculiarities of each real estate market (e.g. Swiss market is a rental market) 9
Benchmarking Houses: Survival Analysis helps! Case study: Swiss rental market! Landlord typically sets a non-negotiable asking price! Landlord has limited information about the willingness to pay of potential tenants! She faces the following trade-off: a higher list price will raise the expected sale rent but, at the same time, it will reduce the number of potential tenants, increasing the Time On Market (TOM) 10
Survival Analysis: Why do we use it?! Statistical problem: At any given time, many offers are still open (censoring)! Use survival analysis to solve the censoring problem! Hint: Follow Paul D. Allison's book 'Survival Analysis Using the SAS System' 11
Step 1: Know Thy Data! Which factors could affect Time on Market?! Use PROC LIFETEST as an explorative tool Proportion of apartments open for rent 100% 75% 50% 25% 0% Zurich! Location? Switzerland 0 25 50 75 100 125 150 175 Duration [days] 12
Step 2: Compute the Estimated Market Rent! Estimated market rent is the rent that an apartment of given size, location, quality, age etc. is expected to achieve on the market, on average! Use hedonic model to estimate the market rent! Compute the difference between the observed list rent and this estimate (= hedonic residual)! This difference carries all the information on whether an offer is too expensive or is a bargain 13
Step 3: Use Proc PHREG to Model TOM! Model survival time (TOM) as a function of a baseline hazard and explanatory variables: Hedonic residual Entry date Location...! proc phreg data=survival outest=parameter ; model duration*completed(0)= hedonic_residual listed_jan listed_feb <...> city1 city2 <more covariates> / ties=efron; baseline out=base survival=s ; run; 14
Step 4: Check for Functional Model Adequacy! Is the Proportional Hazard model justified? Log Cumulative Hazard Rate 0.5-0.5-1.5-2.5-3.5 City 1 City 3 City 2-4.5 0 10 20 30 40 50 60 70 80 Duration [days]! No real need for more advanced techniques, like informative truncation models, NN and the like 15
Step 5: Check for Model Fit! Out-of-sample performance! Gain chart: Order properties according to their probability of being rent and compare this to a random ordering Percentage of completed events 100 80 60 40 20 0 1 2 3 4 5 6 7 8 9 10 Modeled Random Decile 16
Step 6: Results! People clearly react to the over- or underpricing of apartments, i.e. they have a precise guess about correct rents 1 Impact on hazard rate 0.8 0.6 0.4 0.2 0-0.2-0.4-0.6-0.8-1 overpriced -45% -30% -15% 0% 15% 30% 45% longer duration Property is under-/overpriced 17
The bottom line! Help in choosing an optimal price strategy! Tool for active market segmentation as some offers are clearly more valuable than others to the potential tenants! Illustrates opportunities of data-driven approach to mortgage and real estate markets 18