Popular in Economics

Banks and Policy Organisations

Macroeconomic modelling of CBDC: a critical review

European Central Bank Ulrich Bindseil

Banks and Policy Organisations

Financing Private Credit

Federal Reserve Bank of New York Nina Boyarchenko

Economics

Failing Banks

Massachusetts Institute of Technology Emil Verner

Economic History

Golden Fetters

UC Berkeley Barry Eichengreen

Economic Inequalities

Visions of Inequality

City University of New York Branko Milanovic

Economics

School Finance Equalization Increases Intergenerational Mobility

Yale University Barbara Biasi

Economics

Olivier Blanchard on U.S. Pandemic-Era Inflation

Peterson Institute for International Economics Olivier Blanchard

Efficient Likelihood Evaluation of State-Space Representations

View more insights in Social Sciences - Economics - Econometrics

View more figures from University of Pittsburgh

View the page of David DeJong

Read the Study

Next in Econometrics

The Cross-Sectional Implications of the Social Discount Rate

Maya Eden | Brandeis University

Noisy Measurements Are Important, the Design of Census Products Is Much More Important

John Abowd | Cornell University

How can the evaluation of state-space integrals be made more efficient? “Efficient Likelihood Evaluation of State-Space Representations” introduces the EIS filter, building on previous particle filter methods, to provide continuous and adapted approximations that enhance the computation of these integrals, particularly in applications like DSGE models.

Image courtesy of interviewee. November 21, 2024

Log-in or Sign-up to Faculti
Currently viewing this subject insight as a guest. You have insight(s) remaining for this month. Login to view 8000+ figures on the platform.

Generating Workspace...

Export as PDF

This video discusses numerical techniques used to calculate integrals that cannot be calculated analytically, with a focus on Bayesian inference and importance sampling. When direct sampling from the posterior distribution is not feasible, importance sampling can provide estimates by introducing a sampling density that draws parameters which are then weighted according to their relevance to the posterior distribution. This video method is mentioned in the context of calculating expectations and allows for approximating characteristics of complex models by generating samples.

This video also introduces the concept of filtering within dynamic models, where observable and unobservable state variables interact. Filtering helps infer the unobservable states based on observed data over time. The limitations of traditional filters, such as particle filters, are highlighted, noting their inefficiency particularly when the observed data suggests a certain state distribution that does not align well with the underlying model.

Lastly, the efficient importance sampler is proposed as a solution to these inefficiencies by combining these distributions to refine the sampling process. This video strategy leverages linear normal approximations to hone in on the correct parameter space, broadening the scope of models that can be effectively analyzed while minimizing numerical error. Efficient inference allows for more complex models to be evaluated feasibly by reducing variations that arise from numerical approximations.

Language:

Export as PDF

Export as PDF
Here are five key points from the video:

1. **Need for Numerical Techniques**: Numerical techniques are essential for approximating integrals and characteristics of posterior distributions in Bayesian inference when analytical solutions are not feasible. These techniques, such as importance sampling, allow for approximating complicated functions of model parameters.

2. **Importance Sampling Methodology**: Importance sampling is a technique used when the posterior distribution cannot be directly sampled. It involves calculating a weighted average of function values derived from parameter draws, where the weights account for the ratio between the posterior distribution and the importance sample distribution, thus ensuring that the approximation remains unbiased.

3. **Influence of Numerical Error**: The accuracy of numerical estimates diminishes with numerical error, which can arise from the random draws in sampling processes. Increasing the number of draws can reduce numerical error, while the efficiency and alignment of the sampling distribution with the posterior distribution significantly impact the overall estimation error.

4. **Dynamic Models and Filtering**: In the convideo of dynamic models that evolve over time, filtering is necessary for making inferences about unobserved state variables based on observed data. The challenge lies in integrating over possible realizations of these states, especially when dealing with non-linear and non-normal distributions.

5. **Advancements through Particle Filtering**: Particle filtering techniques help manage the inefficiencies in traditional methods by using a swarm of particles to estimate state probabilities over time. By combining linear normal approximations of sampling densities with target distributions, it becomes possible to navigate more complex models beyond the constraints of linear and normal assumptions, enhancing the scope of viable analytical models.

Export as PDF
Here are some key quotes extracted from the video:

1. **On the necessity of numerical techniques:**
- "Numerical techniques arise, given the need to calculate integrals that cannot be calculated analytically, but rather need to be approximated."

2. **Regarding Bayesian inference:**
- "When I can't calculate the characteristics of the posterior distribution of those functions analytically, I can resort to numerical techniques to do so."

3. **On sampling from the posterior distribution:**
- "In that case, I can approximate the intervals that I'm after by obtaining random draws from that normal distribution."

4. **On the law of large numbers:**
- "By the law of large numbers, the estimates that I get will be arbitrarily close to the true value."

5. **On importance sampling:**
- "Importance sampling is a tool that can be used when the posterior distribution itself is not available as a random number generator."

6. **On the efficiency of the importance sampler:**
- "The quality of the sampler that we use has a huge impact on the numerical error."

7. **On filtering and state variables:**
- "Filtering is to make inferences about the value of the state period by period."

8. **On the limitations of linear and normal assumptions:**
- "The shortcoming of that is the linearity and the normality."

9. **On particle filtering:**
- "In a particle filtering scenario, we've made an inference about the value of the state in period T minus one."

10. **On the challenges of particle filtering:**
- "The problem can arise in that setting... the suggested values of the states in time T aren't well aligned with what the model would tell us."

11. **On combining distributions for efficiency:**
- "What the efficient importance sampler does then is it takes these two distributions and combines them."

12. **On the expansion of models that can be evaluated:**
- "This unleashes the shackles of the linear normal assumption... the universe of models that we can look at are greatly expanded."

These quotes capture the essence of the discussion around numerical techniques, Bayesian inference, importance sampling, filtering in dynamic models, and the challenges of working with non-linear models.

X-TIMESTAMP-MAP=LOCAL:00:00:00.000,MPEGTS:90000

1
00:00:00.245 --> 00:00:03.925
Numerical techniques, arise, given the need

2
00:00:03.925 --> 00:00:08.245
to calculate integrals that, cannot be,

3
00:00:08.245 --> 00:00:11.805
calculated analytically, but rather need to be approximated.

4
00:00:12.545 --> 00:00:17.285
So as a lead example, is, is Bayesian inference of

5
00:00:18.315 --> 00:00:19.685
complicated functions

6
00:00:19.765 --> 00:00:21.805
of the underlying parameters of a model.

7
00:00:22.755 --> 00:00:26.135
Um, when I can't calculate the characteristics

8
00:00:26.135 --> 00:00:28.375
of the posterior distribution of those functions

9
00:00:28.935 --> 00:00:33.535
analytically, I can resort to numerical techniques to do so.

10
00:00:34.155 --> 00:00:36.655
As a, as a simple, lead example.

11
00:00:37.075 --> 00:00:39.615
In this context, suppose

12
00:00:39.645 --> 00:00:44.055
that I can use the posterior distribution itself to generate

13
00:00:44.975 --> 00:00:47.575
artificial realizations of model parameters.

14
00:00:47.915 --> 00:00:49.975
So, for example, suppose the posterior distribution

15
00:00:49.975 --> 00:00:51.135
is a normal distribution.

16
00:00:52.035 --> 00:00:54.935
In that case, I can approximate the intervals that I'm

17
00:00:54.935 --> 00:00:58.955
after by obtaining random draws from

18
00:00:58.955 --> 00:01:00.155
that normal distribution.

19
00:01:00.855 --> 00:01:02.315
For each draw of the parameters,

20
00:01:02.425 --> 00:01:04.555
calculate the underlying function of interest,

21
00:01:05.345 --> 00:01:09.005
and then simply calculate the sample average of the,

22
00:01:09.005 --> 00:01:11.645
of those, function values.

23
00:01:12.505 --> 00:01:15.685
And by the law of large numbers, the estimates

24
00:01:15.685 --> 00:01:18.445
that I get will be arbitrarily close to the true value.

25
00:01:18.475 --> 00:01:21.445
Important sampling is a tool that can be used

26
00:01:22.235 --> 00:01:26.125
when the posterior distribution itself is not available

27
00:01:26.385 --> 00:01:29.965
as a random number generator, which is typically the case.

28
00:01:30.705 --> 00:01:31.885
So instead of being able

29
00:01:31.885 --> 00:01:35.165
to draw parameters from the posterior distribution,

30
00:01:36.085 --> 00:01:40.545
I introduce an important sampling density, which needs

31
00:01:40.545 --> 00:01:41.785
to have several characteristics.

32
00:01:41.805 --> 00:01:44.705
But one is that I can obtain parameter draws

33
00:01:45.215 --> 00:01:46.345
from that distribution.

34
00:01:47.355 --> 00:01:49.255
In that case, just like in that simple case,

35
00:01:49.715 --> 00:01:51.455
for every parameter that I draw,

36
00:01:52.135 --> 00:01:53.855
I calculate the function of interest.

37
00:01:55.495 --> 00:01:57.795
But in this case, instead of a simple sample average,

38
00:01:58.495 --> 00:02:01.555
I'm gonna calculate a weighted average where the weight

39
00:02:02.255 --> 00:02:05.515
is determined by the ratio of the posterior distribution

40
00:02:05.735 --> 00:02:07.435
of the parameters, which is important

41
00:02:07.595 --> 00:02:10.675
'cause it tells me the post posterior probability

42
00:02:10.815 --> 00:02:14.275
of any particular parameterization divided

43
00:02:14.375 --> 00:02:17.955
by the important sampler itself evaluated at that drawing.

44
00:02:18.735 --> 00:02:21.035
That's important because the important

45
00:02:21.105 --> 00:02:22.275
sampler is just a tool.

46
00:02:23.175 --> 00:02:26.355
It should not influence the answer that I get

47
00:02:26.815 --> 00:02:28.515
to my, my calculation.

48
00:02:28.515 --> 00:02:30.315
If you write down a different distribution than me,

49
00:02:30.335 --> 00:02:32.075
we still need to get the same answer.

50
00:02:33.075 --> 00:02:37.015
So by calculating that weighted average, the, the value

51
00:02:37.015 --> 00:02:40.815
of the importance sampler in the denominator offsets the

52
00:02:40.815 --> 00:02:43.975
direct influence that the sampler has in generating

53
00:02:44.055 --> 00:02:45.095
a drawing in the first place.

54
00:02:45.245 --> 00:02:46.975
It's the same idea. At the end of the day,

55
00:02:46.975 --> 00:02:49.855
I've got a numerical approximation for my target.

56
00:02:51.035 --> 00:02:53.855
Now I've just used a tool that gets around the problem

57
00:02:54.475 --> 00:02:56.295
of not being able to draw from the

58
00:02:56.295 --> 00:02:57.935
posterior here's efficiency.

59
00:02:58.675 --> 00:03:03.645
At the end of the day, we won't get the exact answer due

60
00:03:03.645 --> 00:03:05.725
to nu numerical approximation.

61
00:03:06.585 --> 00:03:08.165
So in that first example, if you

62
00:03:08.165 --> 00:03:11.045
and I both use the posterior distribution

63
00:03:11.045 --> 00:03:12.685
to obtain sample drawings, we go home

64
00:03:12.685 --> 00:03:15.765
and we generate 800 million draws.

65
00:03:16.825 --> 00:03:18.855
We're gonna get, we're each gonna get an answer,

66
00:03:19.735 --> 00:03:21.595
and our answers will slightly differ.

67
00:03:22.335 --> 00:03:23.915
And the reason for that difference is

68
00:03:24.015 --> 00:03:25.635
simply numerical error.

69
00:03:26.895 --> 00:03:31.095
Um, the more draws we get in that first example,

70
00:03:31.715 --> 00:03:33.775
the smaller will be the numer numerical error.

71
00:03:33.795 --> 00:03:36.215
And in principle, we can get our answers

72
00:03:36.555 --> 00:03:38.295
as close together as possible.

73
00:03:38.315 --> 00:03:41.455
We can drive that numerical error almost to zero.

74
00:03:41.555 --> 00:03:42.935
We can't ever eliminate it,

75
00:03:42.935 --> 00:03:46.895
but we can get it arbitrarily small with important sampling.

76
00:03:47.605 --> 00:03:51.925
Same idea, except that the quality of the sampler

77
00:03:51.925 --> 00:03:55.865
that we use has a huge impact on

78
00:03:56.725 --> 00:03:58.785
the, that numerical error.

79
00:04:00.055 --> 00:04:03.675
So in the extreme case in which I can write down an

80
00:04:03.675 --> 00:04:08.405
important sampler that perfectly mimics the,

81
00:04:08.505 --> 00:04:11.085
the actual posterior distribution, so

82
00:04:11.085 --> 00:04:14.445
that the relative weights are always identically one,

83
00:04:15.555 --> 00:04:18.205
then I've gotten to the exact context

84
00:04:18.465 --> 00:04:21.325
of essentially being able to draw right from the posterior.

85
00:04:22.435 --> 00:04:26.445
But if the two distributions are not well aligned, then

86
00:04:27.215 --> 00:04:28.845
we'll get very different weights.

87
00:04:28.995 --> 00:04:32.925
Draw by draw. Many of the weights that I, that many

88
00:04:32.925 --> 00:04:35.405
of the draws I obtained from the important sampler just

89
00:04:35.405 --> 00:04:38.395
won't get any importance at all from the

90
00:04:38.395 --> 00:04:39.555
posterior distribution.

91
00:04:40.215 --> 00:04:43.635
And that induces a much higher level of numerical error

92
00:04:44.465 --> 00:04:45.475
into our estimates.

93
00:04:46.125 --> 00:04:49.585
So then the idea behind efficiency in the important sampler

94
00:04:49.585 --> 00:04:53.775
is to tailor the importance sampler as closely

95
00:04:53.915 --> 00:04:56.295
as possible to the posterior.

96
00:04:56.385 --> 00:04:57.935
Let's talk about filtering for a second.

97
00:04:59.685 --> 00:05:02.905
The, the context here is we're working with dynamic models,

98
00:05:02.925 --> 00:05:04.465
models that evolve over time,

99
00:05:05.285 --> 00:05:07.825
and they involve the interface of two different types

100
00:05:07.925 --> 00:05:08.985
of variables.

101
00:05:10.065 --> 00:05:13.925
Variables that I can observe directly, like in the context

102
00:05:13.985 --> 00:05:15.285
of a macro application.

103
00:05:15.565 --> 00:05:20.205
I can observe output consumption, investment, labor hours,

104
00:05:21.435 --> 00:05:23.855
but there are other variables that I can't observe directly,

105
00:05:24.245 --> 00:05:25.815
such as a productivity shock.

106
00:05:26.425 --> 00:05:31.005
Things that cause stochastic wiggling beyond what I can see

107
00:05:31.505 --> 00:05:33.485
for a reason beyond what I can see in the data.

108
00:05:34.815 --> 00:05:37.475
Um, so we've got an interface of variables

109
00:05:37.475 --> 00:05:38.595
that can be observed

110
00:05:38.655 --> 00:05:41.275
and variables that cannot, the variables

111
00:05:41.275 --> 00:05:43.035
that cannot are called state variables,

112
00:05:43.735 --> 00:05:46.675
and they evolve from one period to the next in, in ways

113
00:05:46.675 --> 00:05:47.795
that the model would tell us,

114
00:05:47.935 --> 00:05:49.915
but we can't directly observe them.

115
00:05:51.055 --> 00:05:53.065
When I make inferences about

116
00:05:53.925 --> 00:05:56.785
how the data reflect on the model, the likelihood

117
00:05:56.785 --> 00:06:00.295
of the model given the data, the inferences that I make

118
00:06:01.245 --> 00:06:05.105
cannot depend on a particular realization of the state

119
00:06:05.575 --> 00:06:06.825
because that's unknown.

120
00:06:07.695 --> 00:06:11.155
So what I have to do is to make inferences about the value

121
00:06:11.155 --> 00:06:13.395
of the state period by period.

122
00:06:13.395 --> 00:06:14.435
And that's called filtering.

123
00:06:15.595 --> 00:06:17.805
When I, when I, when I have an understanding

124
00:06:17.905 --> 00:06:21.765
of the probability of the state at any given value

125
00:06:21.765 --> 00:06:23.805
of the state over every period of the sample,

126
00:06:24.555 --> 00:06:26.605
then given those probabilities,

127
00:06:26.885 --> 00:06:29.685
I can make inferences about the model itself.

128
00:06:29.865 --> 00:06:31.205
In state-based evaluations,

129
00:06:31.585 --> 00:06:33.365
I'm making inferences about the states

130
00:06:34.725 --> 00:06:38.105
and about the, the likelihood function of the model.

131
00:06:38.165 --> 00:06:40.865
And both of those then entail integrating over

132
00:06:41.585 --> 00:06:42.905
possible realizations of the state.

133
00:06:43.165 --> 00:06:47.505
In the case in which variables are normally distributed

134
00:06:47.765 --> 00:06:50.555
and the models are linear, we can,

135
00:06:50.775 --> 00:06:52.475
we can do likelihood calculations

136
00:06:52.495 --> 00:06:54.115
and filtering analytically,

137
00:06:54.985 --> 00:06:57.765
and we can get billions of evaluations

138
00:06:57.765 --> 00:06:59.605
of the likelihood function per second.

139
00:07:00.345 --> 00:07:02.685
Um, and so we can very, very quickly

140
00:07:03.515 --> 00:07:05.445
make inferences about the likelihood function.

141
00:07:07.185 --> 00:07:09.565
Um, the shortcoming of that is the linearity

142
00:07:10.105 --> 00:07:11.365
and the normality.

143
00:07:11.705 --> 00:07:15.845
So for example, if I'm interested in models on in which the,

144
00:07:15.945 --> 00:07:18.965
the non-linearity is itself, what is interesting,

145
00:07:19.625 --> 00:07:23.125
for example, in macro models where it's costly

146
00:07:23.145 --> 00:07:26.325
to adjust production technologies, for example,

147
00:07:26.385 --> 00:07:30.605
to scale up a production technology, the cost of doing

148
00:07:30.605 --> 00:07:31.725
that would be exponential

149
00:07:32.025 --> 00:07:35.405
or shifting my investment from one sector to another.

150
00:07:36.395 --> 00:07:39.505
Those non-linearities are inherently important in

151
00:07:39.505 --> 00:07:40.585
the research that we're doing.

152
00:07:40.765 --> 00:07:44.345
To the extent that observations, we, we get lots

153
00:07:44.345 --> 00:07:47.225
and lots of things that look like outliers from the lens

154
00:07:47.225 --> 00:07:50.425
of a normal distribution, which have very thin tails.

155
00:07:51.355 --> 00:07:53.815
It might be that the normal distribution is not a good

156
00:07:53.815 --> 00:07:57.165
assumption, that we need fatter tailed distributions

157
00:07:57.385 --> 00:08:00.445
to account for surprising results

158
00:08:00.705 --> 00:08:03.605
or, you know, what, what would otherwise look like outliers.

159
00:08:04.185 --> 00:08:07.805
So, so that's the context we wanna break away from that,

160
00:08:07.915 --> 00:08:10.365
that linear normal specification.

161
00:08:10.635 --> 00:08:13.285
What got us involved in efficient, important sampling,

162
00:08:13.535 --> 00:08:17.565
quite frankly, is that the filtering techniques

163
00:08:17.565 --> 00:08:19.205
that were being that,

164
00:08:19.205 --> 00:08:23.255
that were being implemented in these contexts, turn out

165
00:08:23.255 --> 00:08:25.335
to be quite inefficient numerically.

166
00:08:25.595 --> 00:08:28.215
Let me quickly explain the, how the particle filter works

167
00:08:29.315 --> 00:08:31.295
and then talk about why they're inefficient

168
00:08:31.315 --> 00:08:33.175
and then what we do to get around that.

169
00:08:33.835 --> 00:08:36.775
So in a part particle filtering, scenario,

170
00:08:37.625 --> 00:08:41.135
we've made an inference about the value of the state

171
00:08:41.755 --> 00:08:43.135
in period T minus one.

172
00:08:43.675 --> 00:08:47.305
And that inference is in the form of a swarm of particles

173
00:08:47.855 --> 00:08:51.185
that give us a probability of a particular state at time,

174
00:08:51.345 --> 00:08:55.205
T minus one, given the data observed up till time T minus

175
00:08:55.345 --> 00:08:59.715
one, we then use that particle as a type

176
00:08:59.715 --> 00:09:03.315
of important sampler to make inferences in period T.

177
00:09:04.505 --> 00:09:08.245
Um, and the way it works is from that swarm of particles

178
00:09:09.025 --> 00:09:11.765
and a state transition density, we can make

179
00:09:13.125 --> 00:09:15.655
predictions about what the time t values

180
00:09:15.835 --> 00:09:19.495
of the states would be without taking into account

181
00:09:20.085 --> 00:09:23.535
what the ober, the observed varia the observed variables

182
00:09:23.535 --> 00:09:24.935
looked like in period T.

183
00:09:25.675 --> 00:09:27.815
So it's sort of a blind sampler.

184
00:09:29.375 --> 00:09:31.635
The problem two problems can arise in that setting.

185
00:09:32.415 --> 00:09:36.195
One is that because we had potentially an unusual

186
00:09:36.305 --> 00:09:38.755
realization of the observables in time t

187
00:09:40.045 --> 00:09:43.855
that the suggested values of the states in time t aren't,

188
00:09:43.875 --> 00:09:46.175
aren't well aligned with what the model would tell us.

189
00:09:46.395 --> 00:09:48.895
If you think about those weighted averages

190
00:09:48.895 --> 00:09:50.855
that I just talked about, you get a lot of weights

191
00:09:50.855 --> 00:09:54.455
that just don't get much importance from, from the data.

192
00:09:55.835 --> 00:09:59.295
The other, scenario that can arise is that the data

193
00:10:00.155 --> 00:10:02.615
are extremely informative about the values

194
00:10:02.675 --> 00:10:06.055
of the states at period T so that the distribution

195
00:10:06.055 --> 00:10:09.335
that we're trying to find is very, very tightly distributed.

196
00:10:10.195 --> 00:10:13.295
And you're groping around with a big fat swarm of particles

197
00:10:13.295 --> 00:10:14.295
and it's hard to find it.

198
00:10:15.495 --> 00:10:17.215
I I like to use the following analogy.

199
00:10:17.685 --> 00:10:19.975
Suppose we're trying to find a needle in a haystack,

200
00:10:20.765 --> 00:10:23.225
but we don't know which haystack the needle is in.

201
00:10:24.165 --> 00:10:26.665
So we're getting suggestions of haystacks.

202
00:10:27.325 --> 00:10:30.065
For example, I'm from the Midwest in the United States, so

203
00:10:30.725 --> 00:10:32.865
the suggestions are coming from the western part

204
00:10:32.865 --> 00:10:36.295
of Nebraska, but the haystack lives in

205
00:10:36.295 --> 00:10:37.335
the eastern part of Iowa.

206
00:10:38.315 --> 00:10:41.095
And as we get suggestions about these haystacks in

207
00:10:41.255 --> 00:10:43.135
Nebraska, we're wasting our time.

208
00:10:43.705 --> 00:10:46.915
We're getting, you know, we're getting lots of draws

209
00:10:46.945 --> 00:10:49.955
that just don't have any information in them.

210
00:10:50.895 --> 00:10:54.195
And so we, we end up being very, very inefficient.

211
00:10:55.495 --> 00:10:58.715
So what in, what the efficient, important sampler does then

212
00:10:59.255 --> 00:11:02.875
is it takes these two distributions and combines them.

213
00:11:03.995 --> 00:11:05.735
Um, and, and we make, we make,

214
00:11:06.355 --> 00:11:10.565
we linear normal approximations of the two of, of the,

215
00:11:10.665 --> 00:11:11.765
of the sampling density

216
00:11:12.265 --> 00:11:15.235
and the target, which, if they're very far apart,

217
00:11:15.265 --> 00:11:18.605
like one's in the northwest northeast corridor

218
00:11:18.625 --> 00:11:21.085
and one's in the southwest corridor, we know

219
00:11:21.085 --> 00:11:23.165
that their weighted average lives somewhere in the middle

220
00:11:23.705 --> 00:11:25.925
of the parameter space in a two dimensional space.

221
00:11:26.705 --> 00:11:29.335
So what we do is we build a linear,

222
00:11:29.335 --> 00:11:31.215
normal approximation in that middle.

223
00:11:32.045 --> 00:11:34.265
We know the truth is non-linear and non-normal,

224
00:11:34.265 --> 00:11:35.865
but we, that's where we start to look.

225
00:11:36.445 --> 00:11:39.025
And so we're in the neighborhood of the right haystack

226
00:11:39.615 --> 00:11:40.945
when we start looking for that needle.

227
00:11:41.215 --> 00:11:45.005
This unleashes the shackles of the linear normal assumption.

228
00:11:45.865 --> 00:11:49.445
We can tackle much more interesting models,

229
00:11:50.755 --> 00:11:52.735
or much more complicated models may maybe

230
00:11:52.755 --> 00:11:55.005
Not interesting, that's a pejorative, you know, that's a,

231
00:11:56.505 --> 00:11:59.365
but just the, the, the, the, the universe of models

232
00:11:59.365 --> 00:12:02.605
that we can look at, are, are just greatly expanded.

233
00:12:03.265 --> 00:12:04.645
Um, and

234
00:12:05.155 --> 00:12:09.405
because we can, obtain inferences more efficiently,

235
00:12:10.075 --> 00:12:13.445
they're actually feasible to, to evaluate.

236
00:12:14.225 --> 00:12:16.725
Um, in the inefficient case, you know, you

237
00:12:16.725 --> 00:12:18.965
and I could come up with estimates of our models, but,

238
00:12:19.345 --> 00:12:22.245
but if our estimators are inefficient,

239
00:12:22.245 --> 00:12:25.245
they're gonna be very different reflecting a lot

240
00:12:25.245 --> 00:12:26.845
of numerical error.

241
00:12:27.705 --> 00:12:32.005
And so driving the numerical error as small as we can out

242
00:12:32.005 --> 00:12:35.495
of the inference just makes it, more,

243
00:12:36.935 --> 00:12:40.265
more feasible to, to tackle more complicated models.

Approximating Integrals: Bayesian Inference and Numerical Techniques

00:00:00 - Numerical techniques, arise, given the need to calculate integrals that, cannot be, calculated analytically, but rather need to be approximated. So as a lead example, is, is Bayesian inference of complicated functions of the underlying parameters of a model. when I can't calculate the characteristics of the posterior distribution of those functions analytically, I can resort to numerical techniques to do so.

00:00:37 - In this context, suppose that I can use the posterior distribution itself to generate artificial realizations of model parameters. So, for example, suppose the posterior distribution is a normal distribution. In that case, I can approximate the intervals that I'm after by obtaining random draws from that normal distribution. For each draw of the parameters, calculate the underlying function of interest,

00:01:09 - of those, function values. And by the law of large numbers, the estimates that I get will be arbitrarily close to the true value. Important sampling is a tool that can be used when the posterior distribution itself is not available as a random number generator, which is typically the case. So instead of being able to draw parameters from the posterior distribution,

Weighted Averages in Bayesian Parameter Sampling

00:01:40 - to have several characteristics. But one is that I can obtain parameter draws from that distribution. In that case, just like in that simple case, for every parameter that I draw, I calculate the function of interest. But in this case, instead of a simple sample average, I'm gonna calculate a weighted average where the weight is determined by the ratio of the posterior distribution of the parameters, which is important

00:02:10 - of any particular parameterization divided by the important sampler itself evaluated at that drawing. That's important because the important sampler is just a tool. It should not influence the answer that I get to my, my calculation. If you write down a different distribution than me, we still need to get the same answer. So by calculating that weighted average, the, the value of the importance sampler in the denominator offsets the

00:02:44 - a drawing in the first place. It's the same idea. At the end of the day, I've got a numerical approximation for my target. Now I've just used a tool that gets around the problem of not being able to draw from the posterior here's efficiency. At the end of the day, we won't get the exact answer due to nu numerical approximation. So in that first example, if you and I both use the posterior distribution

Minimizing Numerical Error in Sampling Techniques

00:03:12 - and we generate 800 million draws. We're gonna get, we're each gonna get an answer, and our answers will slightly differ. And the reason for that difference is simply numerical error. the more draws we get in that first example, the smaller will be the numer numerical error. And in principle, we can get our answers as close together as possible. We can drive that numerical error almost to zero.

00:03:42 - but we can get it arbitrarily small with important sampling. Same idea, except that the quality of the sampler that we use has a huge impact on the, that numerical error. So in the extreme case in which I can write down an important sampler that perfectly mimics the, the actual posterior distribution, so that the relative weights are always identically one,

00:04:18 - of essentially being able to draw right from the posterior. But if the two distributions are not well aligned, then we'll get very different weights. Draw by draw. Many of the weights that I, that many of the draws I obtained from the important sampler just won't get any importance at all from the posterior distribution. And that induces a much higher level of numerical error into our estimates.

Dynamic Models and State Filtering in Bayesian Inference

00:04:49 - is to tailor the importance sampler as closely as possible to the posterior. Let's talk about filtering for a second. The, the context here is we're working with dynamic models, models that evolve over time, and they involve the interface of two different types of variables. Variables that I can observe directly, like in the context of a macro application. I can observe output consumption, investment, labor hours,

00:05:24 - such as a productivity shock. Things that cause stochastic wiggling beyond what I can see for a reason beyond what I can see in the data. so we've got an interface of variables that can be observed and variables that cannot, the variables that cannot are called state variables, and they evolve from one period to the next in, in ways that the model would tell us, but we can't directly observe them. When I make inferences about

00:05:56 - of the model given the data, the inferences that I make cannot depend on a particular realization of the state because that's unknown. So what I have to do is to make inferences about the value of the state period by period. And that's called filtering. When I, when I, when I have an understanding of the probability of the state at any given value of the state over every period of the sample,

Challenges in Linear Normal Models and Non-linearity in Inference

00:06:26 - I can make inferences about the model itself. In state-based evaluations, I'm making inferences about the states and about the, the likelihood function of the model. And both of those then entail integrating over possible realizations of the state. In the case in which variables are normally distributed and the models are linear, we can, we can do likelihood calculations and filtering analytically,

00:06:57 - of the likelihood function per second. and so we can very, very quickly make inferences about the likelihood function. the shortcoming of that is the linearity and the normality. So for example, if I'm interested in models on in which the, the non-linearity is itself, what is interesting, for example, in macro models where it's costly to adjust production technologies, for example,

00:07:30 - that would be exponential or shifting my investment from one sector to another. Those non-linearities are inherently important in the research that we're doing. To the extent that observations, we, we get lots and lots of things that look like outliers from the lens of a normal distribution, which have very thin tails. It might be that the normal distribution is not a good assumption, that we need fatter tailed distributions

Enhancing Filtering Techniques through Efficient Importance Sampling

00:08:00 - or, you know, what, what would otherwise look like outliers. So, so that's the context we wanna break away from that, that linear normal specification. What got us involved in efficient, important sampling, quite frankly, is that the filtering techniques that were being that, that were being implemented in these contexts, turn out to be quite inefficient numerically. Let me quickly explain the, how the particle filter works

00:08:31 - and then what we do to get around that. So in a part particle filtering, scenario, we've made an inference about the value of the state in period T minus one. And that inference is in the form of a swarm of particles that give us a probability of a particular state at time, T minus one, given the data observed up till time T minus one, we then use that particle as a type

00:09:04 - and the way it works is from that swarm of particles and a state transition density, we can make predictions about what the time t values of the states would be without taking into account what the ober, the observed varia the observed variables looked like in period T. So it's sort of a blind sampler. The problem two problems can arise in that setting.

Challenges in State Variable Estimation and Efficient Sampling Techniques

00:09:36 - realization of the observables in time t that the suggested values of the states in time t aren't, aren't well aligned with what the model would tell us. If you think about those weighted averages that I just talked about, you get a lot of weights that just don't get much importance from, from the data. The other, scenario that can arise is that the data are extremely informative about the values

00:10:06 - that we're trying to find is very, very tightly distributed. And you're groping around with a big fat swarm of particles and it's hard to find it. I I like to use the following analogy. Suppose we're trying to find a needle in a haystack, but we don't know which haystack the needle is in. So we're getting suggestions of haystacks. For example, I'm from the Midwest in the United States, so the suggestions are coming from the western part

00:10:36 - the eastern part of Iowa. And as we get suggestions about these haystacks in Nebraska, we're wasting our time. We're getting, you know, we're getting lots of draws that just don't have any information in them. And so we, we end up being very, very inefficient. So what in, what the efficient, important sampler does then is it takes these two distributions and combines them.

Overcoming Linear Constraints in Model Estimation

00:11:06 - we linear normal approximations of the two of, of the, of the sampling density and the target, which, if they're very far apart, like one's in the northwest northeast corridor and one's in the southwest corridor, we know that their weighted average lives somewhere in the middle of the parameter space in a two dimensional space. So what we do is we build a linear, normal approximation in that middle. We know the truth is non-linear and non-normal,

00:11:36 - And so we're in the neighborhood of the right haystack when we start looking for that needle. This unleashes the shackles of the linear normal assumption. We can tackle much more interesting models, or much more complicated models may maybe Not interesting, that's a pejorative, you know, that's a, but just the, the, the, the, the universe of models that we can look at, are, are just greatly expanded. and

00:12:10 - they're actually feasible to, to evaluate. in the inefficient case, you know, you and I could come up with estimates of our models, but, but if our estimators are inefficient, they're gonna be very different reflecting a lot of numerical error. And so driving the numerical error as small as we can out of the inference just makes it, more,