Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2...
Transcript of Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2...
![Page 1: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/1.jpg)
Higher-Order Perturbation& Penalty Functions
Wouter J. Den Haan with contributions by Joris de Wind &Ken Judd
August 16, 2016
![Page 2: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/2.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
Outline
1 Reasons why nonlinearities matter more when modellingidiosyncratic risk
2 Problems with higher-order perturbation solutions
3 Using penalty functions instead of inequality constraints
![Page 3: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/3.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
Non-linearities more important forindividual
Reasons:
1 Higher variance state variables
2 Frictions
3 Inequality constraints matter
![Page 4: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/4.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
Need for higher-order perturbationsolutions?
• for risk to matter =⇒ need at least 2nd-order• welfare comparison =⇒ need at least 2nd-order• for risk premiums to be cyclical =⇒ need at least 3th-order• idiosyncratic risk =⇒ need at least ?th-order• models with interesting frictions =⇒ need at least ?th-order• models about the financial crisis =⇒ need at least ?th-order
![Page 5: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/5.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
Problems of higer-order perturbation
• Well-known problem for lots of model solvers• Higher-order perturbation solutions are often explosive• Standard solution is pruning:
• this creates an ugly distortion of underlying perturbationsolution
• Perturbation solutions have more problems• for example weird shapes
• What can be done?
![Page 6: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/6.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
Outline
• Polynomial approximations and its problems• Pruning and its problems• Understanding what perturbation is• Understanding the flexibility of perturbation• Some ideas on how to exploit this flexibility
![Page 7: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/7.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
Polinomial approximations
x+1 = h(x) ≈ pN(x; αN)
How to find αN?
• Perturbation, Taylor series expansion around x• Projection method
![Page 8: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/8.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
Problems of higher-order polynomials
• oscillating patterns =⇒ not shape preserving• often explosive behavior
x+1 = h(x) ≈ pN(x)
limx→∞
∂pN(x)∂x
= ±∞
limx→+∞
∂pN(x)∂x
= +∞ =⇒ no global convergence
limx→+∞
∂pN(x)∂x
= −∞ =⇒ function must turn negative
![Page 9: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/9.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
Is convergence guaranteed?
• Projection methods:• even uniform convergence (with Chebyshev nodes)• of course only within the grid
• Taylor series expansion• limited radius of convergence• unless function is analytic
• Huge difference!!!• grid is controlled by model solver• radius of convergence is not
![Page 10: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/10.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
Couple examples
• sometimes you get great global approximations• Sometimes you do not. We will look at
• limited radius of convergence• problems with weird/undesirable shapes• stability problems
![Page 11: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/11.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
Example with simple Taylor expansion
Truth is a polynomial:
f (x) = −690.59+ 3202.4x− 5739.45x2
+4954.2x3 − 2053.6x4 + 327.10x5
defined on [0.7, 2]
![Page 12: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/12.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
0.8 1 1.2 1.4 1.6 1.8 210
5
0
5
10truth and level approximation of order: 1
0.8 1 1.2 1.4 1.6 1.8 2
0
20
40
60
80
100truth and level approximation of order: 2
Figure: Level approximations
![Page 13: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/13.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
0.8 1 1.2 1.4 1.6 1.8 2
0
50
100truth and level approximation of order: 3
0.8 1 1.2 1.4 1.6 1.8 2
300
200
100
0truth and level approximation of order: 4
0.8 1 1.2 1.4 1.6 1.8 210
0
10truth and level approximation of order: 5
Figure: Level approximations continued
![Page 14: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/14.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
Approximation in log levels
Truth is no a polynomial.Think of f (x) as a function of z = log(x). Thus,
f (x) = −690.59+ 3202.4 exp(z)− 5739.45 exp(2z)+4954.2 exp(3z)− 2053.6 exp(4z) + 327.10 exp(5z).
![Page 15: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/15.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
0.8 1 1.2 1.4 1.6 1.8 210
0
10truth and log level approximation of order: 1
0.8 1 1.2 1.4 1.6 1.8 2
0
50
100truth and log level approximation of order: 3
0.8 1 1.2 1.4 1.6 1.8 2100
50
0
truth and log level approximation of order: 5
Figure: Log level approximations
![Page 16: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/16.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
0.8 1 1.2 1.4 1.6 1.8 2100
50
0
truth and log level approximation of order: 7
0.8 1 1.2 1.4 1.6 1.8 220
10
0
10
truth and log level approximation of order: 9
0.8 1 1.2 1.4 1.6 1.8 210
0
10truth and log level approximation of order: 12
Figure: Log level approximations continued
![Page 17: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/17.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
ln(x) & Taylor series expansion
ln(x)− ln(x) ≈xx− 1
2!
(xx
)2
+2!3!
(xx
)3
− 3!4!
(xx
)4
+ · · ·+ (−1)N−1 (N− 1)!N!
(xx
)N
=
xx− 1
2
(xx
)2
+13
(xx
)3
− 14
(xx
)4
+ · · ·+ (−1)N−1 1N
(xx
)N
with x = x− x
For which x can we expect things to go wrong?
![Page 18: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/18.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
ln(x) & Taylor series expansions at x = 1
1 1.5 2 2.5
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
x
ln(x)
1st
2nd
5th
25th
1 1.5 2 2.5
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
x
1st
2nd
ln(x) 1st 2nd 5th 25th
![Page 19: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/19.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
ln(x) &Taylor series expansions at x = 1.51 1.5 2 2.5
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
x
ln(x)
1st
2nd
5th
25th
1 1.5 2 2.5
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
x
1st
2nd
ln(x) 1st 2nd 5th 25th
![Page 20: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/20.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
ln(x) &Taylor series expansions at x = 1.51 1.5 2 2.5
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
x
ln(x)
1st
2nd
5th
25th
1 1.5 2 2.5
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
x
1st
2nd
ln(x) 1st 2nd 5th 25th
![Page 21: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/21.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
Perturbation verus projection
• Projection methods =⇒ uniform convergence within the grid• You control the grid
![Page 22: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/22.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
ln(x) & projection approximation in [0,2]
1 1.5 2 2.5
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
x
ln(x)
1st
2nd
5th
25thU = 2
1 1.5 2 2.5
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
x
1st
2nd
U = 3
ln(x) 1st 2nd 5th 25th
![Page 23: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/23.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
ln(x) & projection approximation in [0,3]1 1.5 2 2.5
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
x
ln(x)
1st
2nd
5th
25thU = 2
1 1.5 2 2.5
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
x
1st
2nd
U = 3
ln(x) 1st 2nd 5th 25th
![Page 24: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/24.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
Problems with preserving shape
h(x) = 0.5xα + 0.5x
• α is an integer =⇒ h(x) is a polynomial• α is odd =⇒ ∂h(x)/∂x > 0
![Page 25: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/25.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
Perturbation approximation & preservingshape
0 0.5 1 1.50
0.5
1
1.5
2
2.5
3
3.5
4
4.5
5
x
10th
2nd
true value 2nd 10th
![Page 26: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/26.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
Projection approximation & preservingshape
0 0.5 1 1.50
0.5
1
1.5
2
2.5
3
3.5
4
4.5
5
x
true value 3 nodes 11 nodes
3 nodes
11 nodes
![Page 27: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/27.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
Problems with preserving shape
• nonlinear finite-order polynomials always have "weird" shapes• weirdness may occur close to or far away from steady state• thus also in the standard growth model
![Page 28: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/28.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
Standard growth model and odd shapesdue to perturbation (log utility)
0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0.55
0.16
0.18
0.2
0.22
0.24
0.26
2nd-order
45 degree
truth
![Page 29: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/29.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
Standard growth model and odd shapesdue to perturbation (log utility)
0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0.55
0.16
0.18
0.2
0.22
0.24
0.26
45 degree
truth
3rd-order
![Page 30: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/30.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
Problems with stability
h(x) = α0 + x+ α1e−α2x
x+1 = h(x) + shock+1
• Unique globally stable fixed point
![Page 31: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/31.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
Perturbation approximation & stability
0 0.5 1 1.5 2 2.5 3 3.5 4
0
0.5
1
1.5
2
2.5
3
3.5
4
x
2nd
true value 2nd 45-degree
x*
x**
![Page 32: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/32.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
Model
maxct,at∞
t=1
E∞
∑t=1
βt−1 c1−νt − 11− ν
− P(at)
s.t.
ct +at
1+ r= at−1 + θt
θt = θ + εt and εt ∼ N(0, σ2)
a0 given.
![Page 33: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/33.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
Penalty function
Standard inequality constraint
a ≥ 0
corresponds to
P(a) =
∞ if a < 00 if a ≥ 0
Flexible alternative:
P(a) =η1η0
exp(−η0a)− η2a.
![Page 34: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/34.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
Our penalty function
• can be approximated globally with Taylor series expansion• linear part, −η2a
• not necessary• steady state can be equal to the one without penalty function
![Page 35: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/35.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
Interpreting the penalty function
1 penalty function implements inequality constraint
• η0 must be very high
2 penalty function is alternative to penalty function
• η0 could be high or low
![Page 36: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/36.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
Calibrating the penalty function
• η0, η1, and η2 can be chosen to match data characteristics• Here:
• different values for curvature parameter, η0• η1 and η2 chosen to match mean and standard deviation of at
• many properties of this model similar to "a ≥ 0" model• but tail behavior is different
![Page 37: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/37.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
FOC
c−νt
1+ r+
∂P(at)
∂at= βEt
[c−ν
t+1]
![Page 38: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/38.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
Penalty term in FOC; eta0=10
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1-0.05
-0.04
-0.03
-0.02
-0.01
0
0.01
0.02
![Page 39: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/39.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
Penalty term in FOC; eta0=20
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1-0.045
-0.04
-0.035
-0.03
-0.025
-0.02
-0.015
-0.01
-0.005
0
0.005
![Page 40: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/40.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
Perturbation solutions when η0 = 10
1 1.5 2 2.5
-0.2
0
0.2
0.4
0.6
0.8
x
a1st
3rd
5th
"Truth" 1st 2nd 3rd 4th 5th
1 1.5 2 2.50
200
400
600
800
1000
1200
![Page 41: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/41.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
Perturbation solutions when η0 = 20
1 1.5 2 2.5
-0.2
0
0.2
0.4
0.6
0.8
x
a 1st
2nd3rd
4th
5th
"Truth" 1st 2nd 3rd 4th 5th
1 1.5 2 2.50
200
400
600
800
1000
1200
![Page 42: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/42.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
Perturbation and higher uncertainty
• oscillations more problematic when σ ↑• (more likely to get into problematic part)
• but higher-order perturbation solution adjust when σ ↑• (problematic part may move away from steady state)
![Page 43: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/43.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
Fifth-order perturbation and uncertainty
1 1.5 2 2.5-0.2
-0.1
0
0.1
0.2
0.3
0.4
0.5
x
a
5th = 0.3
= 0
![Page 44: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/44.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
Simulating
• 2nd & 3rd explode• 4th & 5th are inaccurate
![Page 45: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/45.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
Pruning - procedure
All steady states are set equal to 0 to simplify notation
![Page 46: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/46.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
Pruning - procedure
1. Split up perturbation solution into two partspN,pert(at−1, θt) =
linear part γN,kat−1 + γN,θθt
nonlinear part +pN,pert (at−1, θt)
![Page 47: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/47.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
Pruning - procedure
2. Simulate a∗t using
a∗t = γN,ka∗t−1 + γN,θθt
3. Simulate aprune,t using
aprune,t= γN,kaprune,t−1 + γN,θθt + pN,pert
(a∗t−1, θt
)
![Page 48: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/48.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
Pruning - procedure
aprune,t= γN,kaprune,t−1 + γN,θθt + pN,pert
(a∗t−1, θt
)• aprune,t is not a function of just the state variables
• aprune,t−1 and θt
• aprune,t also depends on a∗t−1 =⇒aprune,t is a correspondence of state variables
![Page 49: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/49.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
Perturbation principle• Objective of perturbation: If h(x) is such that
f (h(x)) = 0 ∀x
then we want to solve for
happrox(x) = h (x) +∂h(x)
∂x
∣∣∣∣x=x
(x− x) +∂2h(x)
∂x2
∣∣∣∣x=x
(x− x)2
2!
+ · · ·+ ∂nh(x)∂xn
∣∣∣∣x=x
(x− x)n
n!
• Pruning does not generate a function of the form
h(x)
• As a function of x you get a correspondence
![Page 50: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/50.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
Why don’t you get a policy function?
Additional state variables introduced by pruning procedure=⇒ hprune is not a function of x
![Page 51: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/51.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
Why don’t you get a policy function?
1 2 3 40.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
0.5
t
a
Modified 1st
2nd
Pruned 2nd
A
B
C
![Page 52: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/52.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
Pruning - graphs
Our model only has one state variable, xt = at−1 + θt
• Generate aprune,tTt=1
• plot aprune,t as function of xprune,t = aprune,t + θt
![Page 53: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/53.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
Pruning - second-order
![Page 54: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/54.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
Pruning - third-order
![Page 55: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/55.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
Pruning - fourth-order
![Page 56: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/56.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
Pruning - fifth-order
![Page 57: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/57.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
Improvements
• simple improvements• improvements based on alternative perturbation solutions
![Page 58: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/58.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
Measuring data
Data:length of observed data set Tnobs :
observed data yTnobs = yt,dataTnobst=1 :
moment of interest M(
yTnobsi
)
![Page 59: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/59.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
Original Kydland and Prescott approach:
Model:data generated in ith replication yTnobs
i = yt,iTnobst=1 :
mean of moment of interest MI =∑I
i=1 M(
yTnobsi
)I
st. dev. of moment of interest∑I
i=1
(M(
yTnobsi
)−MI
)I
![Page 60: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/60.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
Most common approach
Model:data generated in 1 replication yTlarge
i = yt,iTlarget=1 :
mean of moment of interest M(
yTlargei
)st. dev. of moment of interest 0
![Page 61: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/61.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
Differences
• In general:lim
Tlarge→∞M(
yTlargei
)6= lim
I→∞MI
except for first-order moments• KP approach deals with fact that small sample results maydifferent
![Page 62: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/62.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
Back to explosive perturbation solutions
• (perturbation) approximations explode =⇒use KP instead of the Tlarge approach
• But sharply diverging behavior still possible• Solution: simply exclude those replications• Drawbacks:
• need a criterion to exclude• need initial conditions
![Page 63: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/63.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
Exclusion criterion
• M1stI : moment according to first-order perturbation solution
• Exclude ith sample if
M(
yTnobsi
)> ΛM1st
I
• We experimented with Λ = 2, 3
![Page 64: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/64.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
Initial conditions
• Ideally: initial conditions drawn from ergodic distribution• One can approximate this using first-order solution (which isstable)
![Page 65: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/65.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
Understanding perturbation
Let
h(k) = truth
g(k; γ) = approximation
• Find coeffi cients γ such that
∂gn(k; γ)
∂kn
∣∣∣∣x=x
=∂hn(k)
∂kn
∣∣∣∣x=x
for n = 0, 1, · · · , N
![Page 66: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/66.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
Understanding perturbation’s flexibility
1 You are not restricted to use polynomials
2 Values of∂gn(k; γ)
∂kn
∣∣∣∣x=x
for n > N
are not restricted to be anything
![Page 67: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/67.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
Exploiting higher-order degrees of freedom• Suppose you are given
h(k),∂h(k)
∂k,
∂h2(k)∂k
and consider
g(k; η) = η0 + η1(k− k) + η2(k− k)2 + η3(k− k)3
• Standard perturbationη3 = 0
• But this is arbitrary• Derivatives have no information on this• You could use this additional degree of freedom to implementanother desired property
![Page 68: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/68.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
Exploit functional form flexibility• Suppose you are given
∂hn(k)∂kn
∣∣∣∣x=x
for n = 0, 1, · · · , N
• You would like to use
g(k; η) = η0g0(k) + η1g1(k) + · · ·+ ηNgN(k)
• Solve for the values of a from the following N+ 1 equations
∂hn(k)∂kn
∣∣∣∣k=k
=[η0, η1, · · · , ηN
]
∂gn0(k)
∂kn
∣∣∣k=k
...∂gn
N(k)∂kn
∣∣∣k=k
![Page 69: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/69.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
Simple example
1/x
• Fourth-order Taylor series expansion
1/x ≈ 1− (x− 1) + 2 (x− 1)2 − 6 (x− 1)3 + 24 (x− 1)4
• Alternative
1/x ≈ η0e−2(x−1) + η1e−2(x−1)(x− 1) + η2e−2(x−1)(x− 1)2
+η3e−2(x−1)(x− 1)3 + η4e−2(x−1)(x− 1)4
• note that this is not a transformation
![Page 70: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/70.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
Standard Taylor expansion
0.5 1 1.5 2 2.50
20
40
60
80
100
120
![Page 71: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/71.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
Alternative Taylor expansion
0.5 1 1.5 2 2.50.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
![Page 72: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/72.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
Generate stable perturbation solutions
1 Use alternative basis functions
• trivial modification for 2nd-order perturbation
2 Use a perturbation-consistent weighted combination
![Page 73: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/73.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
Alternative basis functions
• Original model:F(k−1, k, k+1) ≡ 0
F (k−1, h(k−1), h(h(k−1))) ≡ 0
• From (say) Dynare you get
g(k; η) = η0 + η1k− k) + η2(k− k
)2
![Page 74: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/74.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
Alternative basis functions
• Instead of g(k; η) use g(k; η)
g(k; η) = η0 + η1(k− k
)+ η2
(k− k
)2 exp(−(k− k
)2)
• Globally stable for |η1| < 1
![Page 75: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/75.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
Alternative basis functions
• Implementing perturbation principle: solve η from
g(k; η) = h(k)
∂g(k; η)
∂k=
∂h(k)∂k
∂g2(k; η)
∂k2 =∂2h(k)
∂k2
• Amazing but true:η = η
![Page 76: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/76.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
Alternative basis functions
0 0.5 1 1.5 2 2.5 3 3.5−1.5
−1
−0.5
0
0.5
1
1.5
2
2.5
3
cash on hand
asse
tsPolicy function assets
1st−order2nd−order2nd−order alternative45−degree (shifted)
![Page 77: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/77.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
Alternative basis functions
• How to remain closer to underlying second-order perturbation?
• Useexp
(−α(k− k
)2)
and choose low value of α
![Page 78: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/78.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
Perturbation consistent weighting
• Original model:F(k−1, k, k+1) ≡ 0
• add new variable y and new equation
k =y × exp−α(k−1 − k)2
+(
η1st,0 + η1st,1k−1
)×
(1− exp−α(k−1 − k)2
)• α controls speed of convergence towards stable part
![Page 79: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/79.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
Perturbation consistent weighting
• Solve for perturbation solutions of hk(k−1) and hy(k−1)
• Do not use hk(k−1), but use
k = hk(k−1) =
hy(k−1) × exp−α(k−1 − k)2
+(
η1st,0 + η1st,1k−1
)×
(1− exp−α(k−1 − k)2
)
![Page 80: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/80.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
Perturbation consistent weighting
• Approximation is a function not a correspondence• Derivatives of hy(k−1) correspond to true derivatives at k =⇒• Derivatives of hk(k−1) correspond to true derivatives at k• and k = hk(k−1) is globally stable
![Page 81: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/81.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
Note the difference with
k = hk(k−1) =
pkth(k−1) × exp−α(k−1 − k)2
+(
η1st,0 + η1st,1k−1
)×
(1− exp−α(k−1 − k)2
)• Derivatives of hk(k−1) are not correct derivatives of h(k−1)
![Page 82: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/82.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
Perturbation consistent weighting
1 1.5 2 2.5 3 3.5
-0.2
0
0.2
0.4
0.6
0.8
x
a
"Truth" 1st 2nd 3rd 4th 5th
1 1.5 2 2.5 3 3.50
200
400
600
800
1000
1200
x
![Page 83: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/83.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
Perturbation consistent weighting
400 420 440 460 480 500−0.2
0
0.2
0.4
0.6
0.8
time
asse
tsSimulation assets
truth5th−order5th−order weighting
![Page 84: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/84.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
Perturbation consistent weighting
400 420 440 460 480 500−0.2
0
0.2
0.4
0.6
0.8
time
asse
tsSimulation assets
truth3rd−order3rd−order weighting
![Page 85: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/85.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
How to choose alpha?
How to choose α?
• Not that diffi cult if you can plot the policy function• Make estimated guess
• e.g., 3 standard deviations away from s, weight on first-ordershould be 0.28
• Try different values for α and use accuracy test (e.g. dynamicEuler equation test)
![Page 86: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/86.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
Perturbation consistent weighting
1 1.5 2 2.5 3 3.5
-0.2
0
0.2
0.4
0.6
0.8
x
a
"Truth" 5th-order = 0.5 = 1 = 2 = 5
1 1.5 2 2.5 3 3.50
200
400
600
800
1000
1200
x
![Page 87: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/87.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
Multi-dimensional problems
• Let s be the N× 1 vector of state variables• Solve first-order solution: k = a1st,0 + a′1st,1s• Calculate Ω, the variance covariance matrix of st
![Page 88: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/88.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
How to choose alpha?
Use
k = hy(x) × exp− α
N (s−1 − s)′Ω−1(s−1 − s)
+(a1st,0 + a1st,1s
)×
(1− exp
− α
N (s−1 − s)′Ω−1(s−1 − s))
or
k =hy(x) × exp
− α
N (s−1 − skth)′Ω−1(s−1 − skth)
+(
akth ,0 + akth ,1s)×
(1− exp
− α
N (s−1 − skth)′Ω−1(s−1 − skth)
)
![Page 89: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/89.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
Multidimensional problems
• Try different values for α and use accuracy test
• e.g. dynamic Euler equation test
![Page 90: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/90.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
Penalty functions
• to approximate inequality constraint• to describe feature in actual economy
![Page 91: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/91.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
Overview
• Example• How to choose parameters• Different from inequality constraint?• Blanchard-Kahn conditions• Functional form
• try to get them analytic• stay in space of perturbation approximation
![Page 92: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/92.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
Example
P(a) =η1η0
exp(−η0a)− η2a.
![Page 93: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/93.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
Calibrating the penalty function
• η0, η1, and η2 can be chosen to match data characteristics
• η0 clearly a key parameter
![Page 94: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/94.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
Penalty versus inequality
• different values for curvature parameter, η0
• η1 and η2 chosen to match mean and standard deviation of at• =⇒ these two properties "correct"• how different is tail behavior when no numerical errors aremade?
![Page 95: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/95.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
Lower tail
We look at
• Amin: minimum value of A attained• Q1: first quintile• D1: first decile• D2: second decile
![Page 96: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/96.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
26
Lower-tail CDF errors
-0.3000
-0.2500
-0.2000
-0.1500
-0.1000
-0.0500
0.0000
0.0500
η0 =
1
η0 =
2
η0 =
5
η0 =
10
η0 =
20
η0 =
30
η0 =
40
η0 =
50
η0 =
60
η0 =
70
η0 =
80
η0 =
90
η0 =
100
η0 =
120
η0 =
150
η0 =
200
η0 =
300
η0 =
500
η0 =
1000
η0 =
2000
η0 =
10000
AminQ1 D1 D2
Figure 4.2 lower-tail CDF errors
In table 4.2a and 4.2b the relative deciles stand out as there seems to be no convergence
to the constraint model. Yet, this is only a matter of numerical difficulties. Small errors in
the lower-tail of the cumulative density function result in large errors in the relative
deciles. This amplification is caused by the steepness of the cumulative density function
around zero, that is the large mass around zero. Even if the errors in the cumulative
density function of the punishment model become very small, the large errors in the
relative deciles will not disappear, because the cumulative density function of the
constraint model is vertical at zero.
In figure 4.3, we present for a few correlations the percentage error in comparison with
the constraint model. It catches the eye that all correlations are too high. We brief on the
correlation that is the least off and the correlation that is the most off. First, the
correlation between consumption and income is the least off. This implies that
consumption smoothing is not too much affected by using a punishment function instead
of the borrowing constraint. The punishment function facilitates consumption smoothing
![Page 97: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/97.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
First-order condition
c−νt
1+ r+ η1 exp
(−η0a
)− η2 = βEt
[c−ν
t+1]
![Page 98: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/98.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
Suppose there is no penalty function
Eigenvalues
λ+ = 1+ r
λ− =1
(1+ r)β
typical impatience assumption:
β <1
1+ r
=⇒BK conditions not satisfied
![Page 99: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/99.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
How to satisfy Blanchard-Kahn conditions?
• Put in penalty function• Will Blanchard-Kahn condition be satisfied?
• possibly not for high value of η0• penalty term too flat at high η0 values
![Page 100: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/100.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
How to satisfy Blanchard-Kahn conditions?
• Are local dynamics necessarily unstable for high η0?• NO
• with uncertainty:• higher-order perturbation change first-order term
• How to implement this with Dynare?
![Page 101: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/101.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
Functional forms used
• Preston and Roca (2007)
P(a) =η
(a− a)2
• Kim, Kollmann, and Kim (2010)
η
(ln
aass− a− ass
ass
)• Drawback of both:
• not analytic
![Page 102: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/102.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
Functional forms used
• Den Haan and De Wind (2010)
P(a) =η1η0
exp(−η0a)− η2a
• Advantage• analytic
• Drawback• not clear how perturbation solution will behave
![Page 103: Higher-Order Perturbation & Penalty Functions · Calibrating the penalty function h 0, h 1, and h 2 can be chosen to match data characteristics Here: di⁄erent values for curvature](https://reader033.fdocuments.us/reader033/viewer/2022042914/5f4d703456084b7d8169b1da/html5/thumbnails/103.jpg)
Introduction problems DSGE model perturbation solution pruning Alternatives to pruning Penalty functions
Possible fix
• Suppose you use second-order approximation• Let P(a) be such that
• ∂P(a)∂a = η0 + η1a+ η2a2
• problematic behavior far enough away from steady state