1 of 51

Housing Markets and Climate Risk: from Miami to Italy

Microdata in Macromodels, WU Vienna, 29/04/2025

Marco Pangallo,

Matteo Coronese, Francesco Lamperti, Guido Cervone, Francesca Chiaromonte, Anna Bellaver, Lorenzo Costantini, Ariadna Fosch, Jacopo Lenti Anna Monticelli, David Scala

​

2 of 51

3 of 51

Research questions

  • To what extent is climate risk priced in the housing market?
  • What’s the role of beliefs? Are beliefs evolving with climate change?

Project 1:

  • Miami
  • 2000-2019
  • Sea level rise

​

Project 2:

  • Italy
  • 2025-2050
  • Riverine floods

​

4 of 51

Of course, we answer these questions through a data-driven agent-based model!

5 of 51

PROJECT 1: MIAMI

6 of 51

Calibration and initialization

Approximate

Bayesian

Computation

7 of 51

Validation: reproduce 2008 crisis

8 of 51

Can we infer how beliefs on sea level rise risk

changed in Miami from 2000 to 2019?

  • We do not observe beliefs: they are the key latent variable of this model
  • But we observe how prices evolve in areas with different sea level rise risk
  • The coefficient on risk of an hedonic regression can be taken as the observable
  • Estimate this coefficient on both real and simulated data, using data assimilation techniques such as particle filters to make model dynamics track real dynamics by estimating beliefs, as the key latent variable

9 of 51

The case of Little Haiti & Miami Beach

10 of 51

The case of Little Haiti & Miami Beach

11 of 51

Gentrification dynamics �with inferred beliefs?

Constant beliefs

12 of 51

Conclusion

  • Through the data-driven ABM, we quantify how buyers’ behavior changed in Miami in the last 10 years: in 2019, buyers are half as likely to bid for a property at risk compared to 2010
  • I think it’s a very nice combination of micro-econometrics and ABM: because ABMs generate micro-data, many more cross-fertilizations are possible
  • Still much to do on understanding gentrification and belief heterogeneity

13 of 51

PROJECT 2: ITALY

14 of 51

Calibration and initialization

Approximate

Bayesian

Computation

15 of 51

Synthetic homes with generative AI

16 of 51

Validation prices (neighborhood level)

17 of 51

How will the penalty for homes at risk

change from 2025 to 2050?

  • Key observation from our econometric paper: it is the repeated exposure to flood risk, not an occasional (even if strong) flood event that penalizes homes at risk.
  • Simulate floods according to multiple climate models. Each flood contributes a little to beliefs, making more unlikely for buyers to bid for homes at risk.
  • Compare change in price to baseline scenario with no change in beliefs

​

18 of 51

19 of 51

20 of 51

21 of 51

Conclusion

  • Through the data-driven ABM, we predict how the price penalty for homes at risk of riverine floods will change over the next 25 years
  • We unpack how this penalty is due to a decrease in price for homes at risk and an increase in price for homes not at risk, due to a shift in demand. Not possible with observational data
  • The spatial features of the model make it possible to study price dynamics at very granular level

22 of 51

Data-driven ABMs are cool!

https://arxiv.org/abs/2412.16591

23 of 51

EXTRA SLIDES

24 of 51

Initialization/calibration

25 of 51

Initialization/calibration

26 of 51

The model matches some �aggregate dynamics (1)

27 of 51

The model matches some �aggregate dynamics (2)

28 of 51

The model matches some �dynamics across places

29 of 51

Data assimilation

  • Data assimilation is a technique originating in weather forecasting to adjust the state of the model so that it’s compatible with the data
  • Models have latent state variables – variables that you cannot observe but influence the dynamics. Inferring the latent state of the model can bring dynamics closer to data
  • Inference using Kalman or particle filters or variational methods
  • «Black-box» methods as they do not need an explicit formulation of the model

Carrassi et al., WIREs climate change, 2018

30 of 51

Latent state estimation �by making an agent-based model learnable

Monti, Pangallo, de Francisci Morales, Bonchi, Scientific Reports, 2023

31 of 51

Relation to other approaches

Aggregate data

Granular data

Statistical models

Theoretical models

General equilibrium models:

  • Bakkensen and Barrage (2021)
  • Baldauf et al. (2020)

Microeconometric models:

  • Hino and Burke (2021)
  • Bernstein et al. (2019)

32 of 51

Relation to other approaches

Aggregate data

Granular data

Statistical models

Theoretical models

General equilibrium models:

  • Bakkensen and Barrage (2021)
  • Baldauf et al. (2020)

Microeconometric models:

  • Hino and Burke (2021)
  • Bernstein et al. (2019)

Data-Driven Agent-Based Models:

  • Filatova et al. (2023)
  • Ghaffarian et al. (2021)
  • Aerts et al. (2018)
  • This work

33 of 51

The agent-based model of the housing market

Notation. Latin: variables. Greek: parameters. i,j: households. x,y = 1,…,M: homes. t= 1,…, T: time

Buyer decisions:

  1. Enter/exit the market
    • After selling their home
    • By price dynamics
    • From long wait
  2. Decide reservation price
    • Mortgage rule
  3. Choose which home to bid
    • Distance
    • Attractiveness
    • Price
    • Climate risk

Seller decisions:

4. Put home on sale

    • From exogenous factors
    • By price dynamics

5. Decide asking price

    • At first, look around for nearby homes and apply a markup
    • Then, decrease the price is sale unsuccessful

​

Matching:

  1. Match buyer and seller
  2. Determine transaction price and update states

34 of 51

The agent-based model of the housing market

Buyer decisions:

  1. Enter/exit the market
    • After selling their home
    • By price dynamics
    • From long wait
  2. Decide reservation price
    • Mortgage rule
  3. Choose which home to bid
    • Distance
    • Attractiveness
    • Price
    • Climate risk

Seller decisions:

4. Put home on sale

    • From exogenous factors
    • By price dynamics

5. Decide asking price

    • At first, look around for nearby homes and apply a markup
    • Then, decrease the price is sale unsuccessful

​

Matching:

  1. Match buyer and seller
  2. Determine transaction price and update states

Notation. Latin: variables. Greek: parameters. i,j: households. x,y = 1,…,M: homes. t= 1,…, T: time

35 of 51

Buyer decisions: enter/exit the market

Number buyers

Fraction homes

Sensitivity to price changes

 

36 of 51

The agent-based model of the housing market

Buyer decisions:

  1. Enter/exit the market
    • After selling their home
    • By price dynamics
    • From long wait
  2. Decide reservation price
    • Mortgage rule
  3. Choose which home to bid
    • Distance
    • Attractiveness
    • Price
    • Climate risk

Seller decisions:

4. Put home on sale

    • From exogenous factors
    • By price dynamics

5. Decide asking price

    • At first, look around for nearby homes and apply a markup
    • Then, decrease the price is sale unsuccessful

​

Matching:

  1. Match buyer and seller
  2. Determine transaction price and update states

Notation. Latin: variables. Greek: parameters. i,j: households. x,y = 1,…,M: homes. t= 1,…, T: time

37 of 51

Buyer decisions: reservation price

Reservation price buyer i

Fraction of income spent on mortgage

Income

Downpayment

Interest term

Interest rate

Mortgage length

38 of 51

The agent-based model of the housing market

Buyer decisions:

  1. Enter/exit the market
    • After selling their home
    • By price dynamics
    • From long wait
  2. Decide reservation price
    • Mortgage rule
  3. Choose which home to bid
    • Distance
    • Attractiveness
    • Price
    • Climate risk

Seller decisions:

4. Put home on sale

    • From exogenous factors
    • By price dynamics

5. Decide asking price

    • At first, look around for nearby homes and apply a markup
    • Then, decrease the price is sale unsuccessful

​

Matching:

  1. Match buyer and seller
  2. Determine transaction price and update states

Notation. Latin: variables. Greek: parameters. i,j: households. x,y = 1,…,M: homes. t= 1,…, T: time

39 of 51

Buyer decisions: choose home

Distance

Attractiveness

Price

Risk

i’s home

Quality

Reservation

price buyer

Asking

price

Belief in

climate change

​

Climate

risk

​

40 of 51

Buyer decisions: choose home

Distance

Attractiveness

Price

Risk

i’s home

Quality

Reservation

price buyer

Asking

price

Belief in

climate change

​

Climate

risk

​

41 of 51

Buyer decisions: choose home

Distance

Attractiveness

Price

Risk

i’s home

Quality

Reservation

price buyer

Asking

price

Belief in

climate change

​

Climate

risk

​

42 of 51

Buyer decisions: choose home

Distance

Attractiveness

Price

Risk

i’s home

Quality

Reservation

price buyer

Asking

price

Belief in

climate change

​

Climate

risk

​

43 of 51

The agent-based model of the housing market

Buyer decisions:

  1. Enter/exit the market
    • After selling their home
    • By price dynamics
    • From long wait
  2. Decide reservation price
    • Mortgage rule
  3. Choose which home to bid
    • Distance
    • Attractiveness
    • Price
    • Climate risk

Seller decisions:

4. Put home on sale

    • From exogenous factors
    • By price dynamics

5. Decide asking price

    • At first, look around for nearby homes and apply a markup
    • Then, decrease the price is sale unsuccessful

​

Matching:

  1. Match buyer and seller
  2. Determine transaction price and update states

Notation. Latin: variables. Greek: parameters. i,j: households. x,y = 1,…,M: homes. t= 1,…, T: time

44 of 51

Seller decisions: put home on sale

Number sellers

Exogenous prob

Sensitivity to price changes

 

45 of 51

The agent-based model of the housing market

Buyer decisions:

  1. Enter/exit the market
    • After selling their home
    • By price dynamics
    • From long wait
  2. Decide reservation price
    • Mortgage rule
  3. Choose which home to bid
    • Distance
    • Attractiveness
    • Price
    • Climate risk

Seller decisions:

4. Put home on sale

    • From exogenous factors
    • By price dynamics

5. Decide asking price

    • At first, look around for nearby homes and apply a markup
    • Then, decrease the price is sale unsuccessful

​

Matching:

  1. Match buyer and seller
  2. Determine transaction price and update states

Notation. Latin: variables. Greek: parameters. i,j: households. x,y = 1,…,M: homes. t= 1,…, T: time

46 of 51

Seller decisions: set asking price

Price dynamics within distance

Asking price

Markup

Hedonic price

47 of 51

The agent-based model of the housing market

Buyer decisions:

  1. Enter/exit the market
    • After selling their home
    • By price dynamics
    • From long wait
  2. Decide reservation price
    • Mortgage rule
  3. Choose which home to bid
    • Distance
    • Attractiveness
    • Price
    • Climate risk

Seller decisions:

4. Put home on sale

    • From exogenous factors
    • By price dynamics

5. Decide asking price

    • At first, look around for nearby homes and apply a markup
    • Then, decrease the price is sale unsuccessful

​

Matching:

  1. Match buyer and seller
  2. Determine transaction price and update states

Notation. Latin: variables. Greek: parameters. i,j: households. x,y = 1,…,M: homes. t= 1,…, T: time

48 of 51

Match buyers and sellers

49 of 51

The hedonic model

Return

50 of 51

Parameter calibration

Return

51 of 51

Evidence from regression