1

An Agent-based Model forAssessing Financial Vulnerabilities

Rick BookstaberOffice of Financial Research

Isaac Newton InstituteSystemic Risk: Models and Mechanisms

August 28, 2014

2

• Established by the Dodd-Frank Act• Independent agency, housed in the Department of Treasury• Tasks are to

– Support the inter-agency Financial Stability Oversight Council– Facilitate analysis of the financial system– Improve the quality of financial data available to policymakers

• No regulatory authority

Background: The Office of Financial Research

3

• Version 1.0: Historical Data – VaR Models

• Version 2.0: Static Scenarios – Stress Tests

• Version 3.0: Dynamic Interaction – Agent-based Models

Risk Management – Versions 1.0 to 3.0

4

Asset-based Fire Sale• Asset (Price) Shock → Forced Sales → Shock to other Assets

=> Cascades + Contagion

Funding-based Fire Sale (Funding Run)• Funding Shock → Forced Sales → Further Funding Reduction

=> Cascades + Contagion

Leverage- and Liquidity-drivenAsset-based Fire Sales ↔ Funding-based Fire Sales

The Problem to Solve: Fire Sale Dynamics

5

What is an Agent-based Model (ABM)

Agents pursue their activities period by period• Agents are heterogeneous• Can use heuristics rather than optimize• Observe and react to the changing environment• Influence one another; interdependent with dynamic interaction

ExampleAnalysis of traffic flows

Bookstaber (2012), Using Agent-Based Models for Analyzing Threats to Financial Stability, OFR Working Paper No. 3.

6

Applications of the Agent-based Model

Detect Vulnerabilities (Pre-Shock)• What are the dynamic, knock-on effects

Weather Service (Post-Shock)• Are we on the hurricane’s path; how bad will it be

Policy Planning and Actions (Pre- and Post-Shock)• Where do we put the emergency shut-off valves; which do we close• When do we provide asset and funding liquidity

Data Needs• How much can things be improved with better data

7

ABM Schematic – Flows Between the Agents

8

The Agents Caught Up in Fire Sales

BANK/DEALER

Prime Brokerage

Finance Desk

Derivatives Desk

CASH PROVIDERS

HEDGE FUNDS

ASSE

T M

ARKE

T

Trading Desk

Treasury

INVE

STO

RS

INSTITUTIONSOTHER BANK/

DEALERS

INSTITUTIONSOTHER BANK/

DEALERS

Flow of CollateralFlow of Funding

Asset-based Fire Sale Funding-based Fire Sale

9

Maturity transformation • Short-term deposits to long-term loans.

Credit transformation• Structured products with tranches of varying credit risk.• Safe money into funding for risky hedge funds.

Collateral transformation• Lower quality collateral to higher quality collateral.

Liquidity transformation • Market making.• Repackaging assets into liquid vehicles, such as ETFs.

Risk transformation • Selling off part of the return distribution via derivatives.• Tranches with varying risk characteristics.

Transformations of Flows in the ABMThe Financial System as a Production Plant

10

A Model Run

The model can have an any number of agents , markets, and iterations.

In this model parameterization we have:• Three Assets: A0, A1, A2• Two Hedge Funds: HF1, HF2 • Two Bank/Dealers: BD1, BD2• One Cash Provider: CP1

• Run over 1000 iterations

HF1, BD1 Portfolio: {A0, A1}HF2, BD2 Portfolio: {A1, A2}

11

Schematic for Looking at the Network Dynamics

• Thickness of links shows cumulative effect.• Color of links shows intensity of effect in the current period.• Amount of node that is colored shows capital, funding, or price

relative to initial value.

12

Schematic for Looking at the Network Dynamics

• Thickness of links shows cumulative effect.• Color of links shows intensity of effect in the current period.• Amount of node that is colored shows capital, funding, or price

relative to initial value.

13

Schematic for Looking at the Network Dynamics

• Thickness of links shows cumulative effect.• Color of links shows intensity of effect in the current period.• Amount of node that is colored shows capital, funding, or price

relative to initial value.

14

Schematic for Looking at the Network Dynamics

• Thickness of links shows cumulative effect.• Color of links shows intensity of effect in the current period.• Amount of node that is colored shows capital, funding, or price

relative to initial value.

15

Schematic for Looking at the Network Dynamics

• Thickness of links shows cumulative effect.• Color of links shows intensity of effect in the current period.• Amount of node that is colored shows capital, funding, or price

relative to initial value.

16

Schematic for Looking at the Network Dynamics

• Thickness of links shows cumulative effect.• Color of links shows intensity of effect in the current period.• Amount of node that is colored shows capital, funding, or price

relative to initial value.

17

• A0 experiences a 15% price shock• BD1 and HF1 hold A0 in their portfolio• CP holds A0 as collateral• The end of the story for the standard stress test

Period 0: The Static Stress – A 15% Price Shock to A0

Period 0

18

• BD1 and HF1 decrease positions in both A0 and A1• This creates a downward cycle for A0 and a drop in A1• It also affects other agents holding A0 or A1• CP1 reduces funding as its collateral value drops

Period 2: Cascade in A0 and Contagion through A1

Period 0 Period 2

19

• The drop in prices ignites a funding-based fire sale through CP1• The dynamic spreads due to credit exposure from BD1 to BD2 • This can lead to difficulty in identifying the source of contagion

Period 4: Credit and Funding Effects

Period 2 Period 4

20

• The fire sale reaches its end. • In this run BD1, HF1, and HF2 have defaulted

Period 6: Collateral and Capital Damaged

Period 4 Period 6

21

Sources of Shock

Asset Market: Price Shock Cash Provider: Funding Shock

Bank/Dealer: Credit Shock Hedge Fund: Redemption Shock

22

Tracking the Propagation of Shocks

Asset Market: Price Shock Cash Provider: Funding Shock

Bank/Dealer: Credit Shock Hedge Fund: Redemption Shock

23

ConclusionWhat Does Version 3.0 Mean at the Firm Level

Dynamic Stress Testing• Are you on the hurricane's path?• Will you become collateral damage

Salvaging VaR• Crisis VaR and the VaR multiplier

Catching Falling Knives• Being a liquidity supplier of last resort

24

ABM and VaR

• The fire sale cascade leads to a downward skew for the capital post-shock.• Red lines are the mean and 5%/95% envelope, 1000 runs of a 15% shock in Asset 1.

0 2 4 6 8 10 12 14 16 18$0

$2

$4

$6

$8

$10

$12

$14

Period Post Asset 1 Shock

Cap

ital

of F

irm

($B

B)

-5-4-3-2-1 0 1 2 3 4 5 6 7 8 9 10-70%-60%-50%-40%-30%-20%-10%

0%10%20%

99.0% 95.0%Mean 5.0%

Period Post Asset 1 ShockCa

pit

al C

han

ge o

f Fir

m

25

Networks and Agent-based Models

Using the “Maps” to Create a Multi-layer Network• Funding Map• Collateral Map• Assets Map

What agents faciliate movement from one layer to another

Using the ABM to Create and Analyze Dynamical Networks• Nodes provide transformations and to respond to the environment• Changes in the size (and existence) of nodes• Links vary in size of flows, and in their effect on the behavior of the

transformations in the nodes

26

Question: Testing the Agent-based Model

Parameter Realism• Do parameter values of real-world agents lead to real-world dynamics

Comparative Statics • Do things move in the right direction, by the right amount, from a reasonable

initial value• Is there common sense consistency

Stylized Facts• Do we see agents and markets behave in the right way

Back Testing• Can we reproduce past events

27

Question: Populating the Model with Data and Rules

Data• Exposures: dominant investment themes, credit• Funding: sources, durability, leverage, collateral• Prices: “big trade” liquidity• Frequency: Exposures and funding build and change slowly• Completeness: More is better; less can still work

Agents’ Rules • Many actions during stress are pre-determined and non-proprietary

28

Cash Provider

Home

29

Asset

Home

30

Derivatives Desk

Home

31

Treasury

Home

32

Prime Broker

Home

33

Finance Desk

Home

34

Hedge Fund / Trading Desk

35

Hedge Fund / Trading Desk

Home