Testing and Improving Commercial Real Estate Market Segmentations with Cluster Analysis and Neural Network Techniques. Franz Fuerst and Gianluca Marcato. Real Estate Fund Management. Fund managers normally start from the sector vs. region dichotomy Asset allocation of a mixed fund - PowerPoint PPT Presentation
Welcome message from author
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Testing and Improving Commercial Real Estate Market Segmentations with Cluster Analysis and Neural Network Techniques
Franz Fuerst and Gianluca Marcato
School of Real Estate & Planning
Real Estate Fund Management
• Fund managers normally start from the sector vs. region dichotomy– Asset allocation of a mixed fund– Selling proposition of a specialised fund
• … but fund managers also consider other characteristics– Indirectly
• E.g. small funds; several small prop few big props– Directly
• E.g. grade A vs. grade B buildings
• … So can we ‘formalise’ this process ?
2
School of Real Estate & Planning
Research Rationale• Review of cluster analysis technique in Romesberg 84
– Used to discover segmentations within specific sectors: residential (Kroll & Smith 89, Bourassa et al 99 and 03, Wilhelmsson 04), offices (Goetzmann & Watcher 95), hotels (Gallagher & Mansour 00)
– Used to look at portfolio construction (Hoesli et al 97) or trading behaviour in housing markets (Piazzesi & Schneider 09)
– Other previous research suggests that a sector and region classification insufficiently explains variations in return (Lee 01, Andrew 03, Devaney 03)
• Objective of this paper: Explore possible segmentations that have higher predictive power
Research Questions• Are “new” factors relevant to explain real estate
returns?– Property size (i.e. small vs. big properties)– Yields (i.e. value vs. growth properties)– Tenant concentration (i.e. small vs. big number of
tenants)– Lease length (i.e. short vs. long lease)
• What are the implications for benchmarking and forecasting real estate returns?– Should we change our normal way of thinking?
4
School of Real Estate & Planning
Implication: Expanding Asset Allocation
BasicAsset
Allocation
Multi-CriteriaAsset
Allocation
5
School of Real Estate & Planning
Results Overview• Benchmarking
– We should be changing our way of thinking– “New” styles / risk factors explain portfolio returns– Property size is the main “new” risk factor– Part of alpha is paying for exposure to these factors
• Forecasting and Segmentation– We should be changing our way of thinking– Individual real estate returns reveal new segmentations– Yield and tenants concentration are the main “new” risk
factors– Ongoing process to be monitored
6
JPM,Forthcoming
School of Real Estate & Planning
Procedure• 2-step Cluster Analysis
– Using either 10 (PAS) or 14 (PAS2) clusters
– Done for all property and types of property (shopping centres, standard retail, office, standard industrial, distribution warehouses)
• Discriminant Analysis to test consistency of clusters over time and to compare IPD PAS Segments with New Clusters (backward testing)
• Neural Network technique to confirm results of cluster and discriminant analyses (backward testing)
• To be done: Discriminant Analysis to confirm consistency between Cluster Analysis and Neural Network procedure
7
School of Real Estate & Planning
Basics of Cluster Analysis• Minimize within cluster distances (homogeneity)
• Maximize between cluster distances (heterogeneity)
8
x1
x2
max
min
School of Real Estate & Planning
Midlands and SW (1)
0
50
100
150TRR
ERV Growth
E. Yield
Cap.ValueLease Length
# Tenants
Large tenants
Small CV, low # tenants, low ERV growth
9
School of Real Estate & Planning
Wales & NW England (2)
10
0
50
100
150TRR
ERV Growth
E. Yield
Cap.ValueLease Length
# Tenants
Large tenants
Slight exposure to # Tenants
School of Real Estate & Planning
Scotland (5)
0
50
100
150
200TRR
ERV Growth
E. Yield
Cap.ValueLease Length
# Tenants
Large tenants
Long leases, high ERV growth, low # tenants
11
School of Real Estate & Planning
Central London (10)
0
50
100
150
200TRR
ERV Growth
E. Yield
Cap.ValueLease Length
# Tenants
Large tenants
High ERV growth
12
School of Real Estate & Planning
UK with Some Concentration
13
School of Real Estate & Planning
Description of 10 Clusters (03-07)No. DESCRIPTION TRR ERVg EY CV
Factors region region region region region region region region region
PAS segm.
PAS segm.
PAS segm.
PAS segm.
PAS segm.
PAS segm.
PAS segm.
PAS segm.
PAS segm.
local auth.
local auth.
local auth.
local auth.
local auth.
local auth.
local auth.
local auth.
local auth.
20
School of Real Estate & Planning
NN: Sum of Squares Error (Model)
21
School of Real Estate & Planning 22
NN: Sum of Squared Errors (Factor)
School of Real Estate & Planning
NN Estimation of Total Returns
(b) Full set of variables – m1(a) Regions and sub-sectors only – m8
23
School of Real Estate & Planning
Conclusions• New segments have higher predictive power
• Returns are more predictable if we include variables on ERV growth, yield, property size, tenant diversification, lease terms and volatility measures
• Seemingly unrelated regions, sectors, properties move together, -> cluster/discriminant analysis and neural networks detect these patterns
• Both segmentations have their strenghts and weaknesses (IPD: easy to understand what each segment represents, Cluster Segments: higher predictive power)