**Staff** **Working** **Paper** **No**. **633**

**Adaptive** **learning** **and** **labour** **market**

**dynamics**

Federico Di Pace, Kaushik Mitra **and** Shoujian Zhang

December 2016

**Staff** **Working** **Paper**s describe research in progress by the author(s) **and** are published to elicit comments **and** to further debate.

Any views expressed are solely those of the author(s) **and** so cannot be taken to represent those of the Bank of Engl**and** or to state

Bank of Engl**and** policy. This paper should therefore not be reported as representing the views of the Bank of Engl**and** or members of

the Monetary Policy Committee, Financial Policy Committee or Prudential Regulation Authority Board.

**Staff** **Working** **Paper** **No**. **633**

**Adaptive** **learning** **and** **labour** **market** **dynamics**

Federico Di Pace, (1) Kaushik Mitra (2) **and** Shoujian Zhang (3)

Abstract

The st**and**ard search **and** matching model with rational expectations is well known to be unable to

generate amplification in unemployment **and** vacancies. We document a new feature it is unable to

replicate: properties of survey forecasts of unemployment in the near term. We present a parsimonious

model with adaptive **learning** **and** simple autoregressive forecasting rules which provide a solution to

both of these problems. Firms choose vacancies by forecasting wages using simple autoregressive

models; they have greater incentive to post vacancies at the time of a positive productivity shock

because of overoptimism about the discounted value of expected profits.

Key words: **Adaptive** **learning**, bounded-rationality, search **and** matching frictions.

JEL classification: E24, E32, J64.

(1) Bank of Engl**and**. Email: federico.dipace@bankofengl**and**.co.uk

(2) University of Birmingham. Email: k.mitra@bham.ac.uk

(3) Addiko Bank.

The views expressed in this paper are those of the authors, **and** not necessarily those of the Bank of Engl**and** or its committees.

We would like to thank participants of the conference on ‘Expectations in dynamic macroeconomic models’ at the Bank of

Finl**and** (Helsinki), the 1st Birkbeck Centre of Applied Macroeconomics conference at Birkbeck College (London), the

20th Annual Computing in Economics **and** Finance Conference at the BI **No**rwegian Business School (Oslo), the

7th Conference of the Centre of Economic Growth **and** Business Cycles at the University of Manchester, the 2014 Asian

Meeting of Econometric Society at the Academia Sinca (Taipei), 46th Annual Conference of the MMF at the University of

Durham **and** seminar participants at the Universities of Exeter, Cardiff, the Technical University of Vienna **and** the Bank of

Engl**and**. We would especially like to thank William Branch, George Evans, Lien Laureys, Bruce Preston **and** Joseph Pearlman

for useful comments **and** suggestions.

Information on the Bank’s working paper series can be found at

www.bankofengl**and**.co.uk/research/Pages/workingpapers/default.aspx

Publications Team, Bank of Engl**and**, Threadneedle Street, London, EC2R 8AH

Telephone +44 (0)20 7601 4030 Fax +44 (0)20 7601 3298 email publications@bankofengl**and**.co.uk

© Bank of Engl**and** 2016

ISSN 1749-9135 (on-line)

1 Introduction

The Diamond-Mortensen-Pissarides (DMP) search **and** matching model has become the

st**and**ard theory of equilibrium unemployment. Given its popularity, one might expect

strong evidence of the model being consistent with key business cycle facts. However,

Shimer (2005) shows that the st**and**ard search **and** matching model, driven by Total

Factor Productivity (TFP) innovations, has a hard time replicating the cyclical behavior

of its central elements; namely the amplification of **labour** **market** variables, such as

unemployment, vacancies **and** the measure of **labour** **market** tightness present in the US

data. Under the common assumption that wages are negotiated through Nash bargaining

every period, wages tend to absorb most of the productivity innovations, generating little

amplification in profits per hire. 1 This is referred to as the unemployment volatility puzzle

in the literature.

We highlight in this paper, for the firsttimetoourknowledge,anotherimportantfeature

of the data that the st**and**ard search **and** matching model under rational expectations

(RE) is unable to replicate. This feature relates to survey forecasts of the unemployment

rate in the near term. One of the central variables in the search **and** matching model

is the unemployment rate, which households **and** firms have to forecast to make consumption

**and** vacancy posting decisions. Figure 1 shows the one **and** four step-ahead

forecast errors in unemployment rates obtained from the Survey of Professional Forecasters

(SPF). **No**te the systematic over/under-prediction of unemployment rates over the

business cycle made in the surveys by professional forecasters. Most importantly, the

forecast errors are positively correlated with the business cycle. **and** increase with the

forecast horizon. Forecast errors in unemployment rates are much more volatile relative

to output over the 4-quarter horizon. For instance, while the one quarter ahead forecast

error in the unemployment rate is 456 times as volatile as output in the US data, this

measure under RE is close to zero.

Since st**and**ard models assume rational expectations, they are inconsistent with the

survey data because agents do not make systematic errors in these models. The search

**and** matching model with RE is unable to match the properties of the forecast errors.

This paper provides a solution to this puzzle by examining the role of expectation formation

for hiring decisions. We replace the Rational Expectation assumption by a set of

simple autoregressive subjective beliefs that have gained recent popularity in the **Adaptive**

Learning (AL) literature. This simple modification enables the model to match the

general features present in the data: it presents a solution to the unemployment volatility

puzzle as well as match the statistical properties of forecast errors in unemployment found

in US data. Our paper is the first one to highlight the role of expectation formation in

1 Mortensen **and** Nagypal (2007) argue that the performance of the st**and**ard model featuring RE

beliefs depends on the variability of per hire rather than on the asusmption about the cyclicality of

wages.

1 **Staff** **Working** **Paper** **No**. **633** December 2016

the study of hiring decisions **and** its capacity to match important features of labor **market**

data.

Figure 1: Forecast errors of unemployment from SPF. **No**tes. Forecast errors are expresed in percentage points

i.e. forecast unemployment rates minus actual unemployment rates for the period 1968Q4-2015Q2. The solid **and** dotted

lines denote the one-step **and** four-step ahead forecast errors respectively. The figure shows the forecast errors increase as

the time horizon of the forecast increases. B**and**s show the NBER recession dates. The forecast errors are negative during

recessions **and** positive during expansions.

We develop a simple search **and** matching model where wages are negotiated period

by period with the assumption of RE being replaced by subjective beliefs as in the AL

literature. Firms **and** households form forecasts of unemployment rates, wages, profits

**and** interest rates up to the infinite future to make consumption **and** hiring decisions using

simple autoregressive models to form expectations. Firms in our model face a dynamic

problem due to long-lasting employment relations modelled through search frictions. We

assume that agents have incomplete knowledge about the structure of the economy in

that they do not know the preferences **and** technologies of other agents in the economy.

Hiring decisions at the firm level depend crucially upon the perception of future profits

per hire.

We find that not all forms of adaptive **learning** provide a solution to the unemployment

volatility puzzle. In particular, when agents form their forecasts based on perceptions that

take the form of the rational expectations equilibrium (REE) solution then the amplification

results are not too dissimilar from RE. However, strikingly, assuming AL results in a

much better fit to the data than assuming RE when agents use small forecasting models

in their **learning**. In this sense, the results are consistent with Slobodyan **and** Wouters

2 **Staff** **Working** **Paper** **No**. **633** December 2016

(2012). 2 The AL models considered are in a sense minimal departures from the RE solution.

As shown the RE solution involves employment, wages, profits **and** interest rates

depending on the (only) state variable employment **and** the technology shock (assumed

to be of autoregressive process of order one i.e. AR(1) process). This RE solution may

inturnbewritteninanalternativewaywherebywages,profits **and** interest rates are

auto-regressive moving average processes i.e. ARMA(2,1) processes while employment is

an autoregressive process of order two; i.e. an AR(2) process. We consider three variants

of small forecasting models motivated by the nature of the RE solution. The first model

we consider is a minimal departure from the RE solution in that agents assume the key

four endogenous variables (employment, wages, profits **and** interest rates) are all AR(2)

processes. The next model assumes these are all AR(1) processes. The final model we

consider is where agents perceive these four variables to evolve as a VAR(1) process: this

is a popular method of forecasting **and** is particularly appealing as it allows for possible interactions

among the key endogenous variables since agents under incomplete knowledge

are unaware of the general equilibrium restrictions among these variables.

We show that all of these models with adaptive **learning** generate much more amplification

in **labour** **market** variables compared to their RE counterpart (to different degrees).

For instance, for the AR(2) model, unemployment **and** vacancies are 7 **and** 10 times more

volatile than the corresponding measure of output, which is in line with US data; these

numbers are in fact 16 **and** 18 times higher than those in the corresponding RE model.

Our model matches well the statistical properties of forecast errors on unemployment

data taken from the SPF. In particular, the volatility **and** cyclical behavior of unemployment

forecast errors generated by these models match the data closely. This finding is in

sharp contrast with the RE model **and** supports the type of perceived beliefs assumed in

the **learning** model.

In the search **and** matching literature, the job creation condition represents the optimal

decision rule for vacancy posting. Firms post vacancies until the expected marginal

cost of posting a vacancy equals the benefits of hiring an additional worker, which can be

expressed in equilibrium as the (infinite) sum of expected future profits generated at the

margin. 3 Agents with incomplete knowledge about the structure of the economy tend to

become optimistic after a positive TFP innovation **and** this in turn leads to more vacancy

creation; the optimism is greater since firms forecast infinite periods ahead, which results

in more vacancy creation **and** greater amplification. This means that the impact effect

of productivity shocks on the present discounted value of profits is large because agents

make systematic forecast errors about the path of future wages. Since firms carry out

infinite horizon forecasts to choose how many vacancies to posts, firms’ discounting of

2 **No**te, however, that we use the infinite horizon **learning** approach advocated by Preston (2005) unlike

the Euler equation **learning** approach of this paper.

3 The solution to the model with RE beliefs is the same regardless of whether the decision rule for

vacancies is specified recursively or as an infinite sum of future profits per hire.

3 **Staff** **Working** **Paper** **No**. **633** December 2016

future marginal products **and** wages turns out to be central.

Incidentally, there are a large number of studies that have attempted to provide solutions

to the unemployment volatility puzzle under the assumption of RE. Two prominent

solutionsmakerelativelysimplemodifications to the st**and**ard search **and** matching model

to generate greater amplification. The first approach proposed by Hall (2005) **and** Shimer

(2005) introduces real wage rigidities. This means that, as wages cannot fully reflect productivity

shifts, there is further incentive for vacancy creation. 4

The second popular

approach by Hagedorn **and** Manovskii (2008) (henceforth, HM) carries out a simple calibration

exercise that sets the value of non-**market** activity close to the value of search of

the worker. 5 These characterisations of the **labour** **market** have been criticised on the following

grounds: a) microeconometric evidence by Pissarides (2009) **and** Haefke, Sonntag,

**and** van Rens (2013) are suggestive that wages for newly hired workers are cyclical **and**

b) Costain **and** Reiter (2008) show that the implied elasticity of unemployment benefits

with respect to the unemployment rate arising from the calibration in HM is implausibly

large relative to the data.

The main assumption in the paper that economic agents engage in “**learning**” behaviorhasbeenincorporatedintomacroeconomictheory**and**usedinawiderangeof

applications, see Evans **and** Honkapohja (2001, 2003, 2006), Bullard **and** Mitra (2002)

**and** Preston (2005, 2006, 2008). The st**and**ard adaptive **learning** approach treats economic

agents like econometricians who estimate forecast rules, updating the parameter

estimates over time as new data become available. Agents update their forecasting model,

their forecasts of future variables **and** re-solve their dynamic optimization problem in order

to make their decisions. In the context of infinite-horizon **learning**, agents solve

dynamic optimization problems. This **learning** approach can be viewed as a version of

the anticipated utility approach formulated by Kreps (1998) used by Eusepi **and** Preston

(2011) **and** Kuang **and** Mitra (2016) within the context of the RBC model. 6

The point that there is an important divergence between the implied expectations in

macroeconomic models with RE **and** the expectations drawn from surveys (eg. SPF) has

4 Menzio (2005), Gertler, Sala, **and** Trigari (2008), Christoffel **and** Kuester (2008), Gertler **and** Trigari

(2009), Blanchard **and** Gali (2010) **and** Hertweck (2013), amongst others, extend this idea to a general

equilibrium setting.

5 The list of solutions to the unemployment volatility puzzle is not exhaustive; see Menzio **and** Shi

(2011), Colciago **and** Rossi (2011), Alves (2012), Quadrini **and** Trigari (2008), Gomes (2011), Reiter

(2007), Guerrieri (2008), Robin (2011), Petrosky-Nadeau (2013) **and** Petrosky-Nadeau **and** Wasmer

(2013).

6 In the anticipated utility approach recommended by Kreps, agents update their forecasts over time

but do not take into account the fact that their forecasting model will be revised in future periods. This

is a bounded rationality approach, since a full Bayesian approach would recognize the uncertainty in

the parameters of the estimated forecasting model. However, as noted by Cogley **and** Sargent (2008),

a full Bayesian approach in macroeconomic settings is typically “too complicated to be implemented,”

**and** thus the anticipated utility approach is an appealing implementation of bounded rationality. An

alternative approach by Adam **and** Marcet (2011) **and** Adam, Marcet, **and** Beutel (2016) derives jointly

the optimal decisions **and** belief updating rules from the utility maximisation problem in the context of

an asset pricing model.

4 **Staff** **Working** **Paper** **No**. **633** December 2016

een made in the context of other models; see e.g. Adam, Marcet, **and** Beutel (2016),

Kuang **and** Mitra (2016), Slobodyan **and** Wouters (2012), Milani (2011) **and** Ormeno **and**

Molnar (2015). These papers show that systematic errors/gaps in variables such as GDP

**and** interest rates are evident in survey forecasts like the SPF.

The remainder of the paper is organised as follows. Section 2 describes the model

environment. Section 3 describes the RE solution of the model together with the **learning**

models considered in the paper. Section 4 presents the results of the model for the baseline

calibration. Section 5 evaluates the quantitative performance of the model in terms of the

forecast errors of unemployment **and** highlights the similarities **and** differences between

our approach **and** wage rigidities under RE beliefs. Section 6 analyses the robustness of

the results to alternative parameterisations **and** values of the gain parameter. Section 7

concludes.

2 Model

We propose a simple model featuring **labour** **market** search **and** matching frictions as

in Mortensen **and** Pissarides (1994) **and** a form of adaptive **learning** following Preston

(2005) **and** Mitra, Evans, **and** Honkapohja (2013). Our model economy is inhabited by

two types of agents: households **and** firms. There is a representative household consisting

of a continuum of workers that search for jobs if unemployed **and** work for firms if

employed. Following Andolfatto (1996), we make the assumption of perfect risk sharing

at the household level so that both employed **and** unemployed members of the household

consume equal amounts. Firms post job vacancies **and** employ workers with a lag so as

to produce final goods using **labour** as the only input of production. Households consume

the final goods supplied by firms. We assume that agents form their expectations

by updating their beliefs as new information becomes available. Agents make infinite

horizon (IH) forecasts about the future path of wages, interest rates, unemployment **and**

profits by running simple autoregressive models in order to make current decisions about

consumption **and** vacancy posting.

2.1 Labour Market

The **labour** **market** is frictional in that, from the perspective of the firm, it is costly to

post vacancies **and**, from the st**and**point of workers, searching for jobs is a time consuming

process. Every period firms create new vacancies, sought by unemployed workers who

are continuously looking for new job opportunities. Following Shimer (2010), we assume

that workers that are matched at time become productive at the beginning of next

period, +1. Worker-firm matches break up at the exogenous rate, ∈ (0 1). The

aggregate number of matches, , depends positively on both the unemployment rate,

5 **Staff** **Working** **Paper** **No**. **633** December 2016

, **and** aggregate vacancies, . We assume that the matching process is guided by a

matching function that exhibits constant returns to scale

( )= ¯ 1−

(1)

where ¯ denotes the level of efficiency in the matching process, the elasticity of the

matching function with respect to aggregate vacancies **and** the unemployment rate is

given by

=1− (2)

We define the measure of **labour** **market** tightness as

= (3)

Due to the assumption of constant returns in the matching technology, we define the job

finding rate as ( ) = ( 1) = ( ) **and** the job filling rate as

( ) = (1 )= ( ) (4)

**No**te that the job finding rate, ( ), is increasing in the measure of **labour** **market**

tightness **and** the job filling rate, ( ), decreasing.

2.2 Households

In our model economy there is a large representative household consisting of a continuum

of members in the unit interval that maximises its life-time utility

H ( )= ∗

∞X

− U ( ) (5)

=

where ∗ denotes the subjective expectation of the household at time . H ( ) denotes

the value function of the household, which depends on last period savings **and** the period

employment rate. The period utility of the representative household depends positively

on consumption, , **and** negatively on employment, , **and** is defined by

U ( )=log −

where 0 is a parameter that captures the disutility of employment at the level of the

household. The flow budget constraint is given by

+1

= + + − (6)

6 **Staff** **Working** **Paper** **No**. **633** December 2016

where denotes household savings at the end of period , the real interest rate **and**

aggregate profits, which are rebated to the household at the end of each period. There

is a continuum of firms indexed by ∈ [0 1]; the aggregate employment rate is the sum

of employment across firms, = R 1

0 , while = R 1

0 **and** = R 1

0

are the pooled wage bill **and** profits respectively. In addition, the employment rate at the

household level evolves according to

+1 =(1− ) + ( ) (7)

The representative household chooses the level of consumption to maximise its life-time

utility, (5), subject to the flow budget constraint, (6), **and** law of motion of employment,

(7). The household’s problem can be written in terms of the following Bellman equation

H ( )= max

+1

{log − + ∗ H ( +1 +1 )}

together with the constraints (6) **and** (7). By combining the first order conditions with

respect to **and** , we obtain the st**and**ard Euler equation

1

= ∗

(8)

+1

This condition states that the household’s marginal utility derived from consumption at

time must equal the marginal utility of consumption derived at time +1expressed in

terms of time .

Theenvelopeconditionwithrespectto gives

H ( )

=

− + [1 − − ( )] ∗

H ( +1 +1 )

+1

· (9)

This condition states that the net marginal value of having a member of the household

employed at firm is equal to the net flow value of employment, the difference between

the wage, expressed in terms of utility, **and** the utility cost of working, plus the net

continuation value of employment.

By iterating forward the budget constraint, (6), **and** substituting for future values of

savings, we can find an expression for consumption

+ ∗

∞X

=1

−1

+ + = + + + ∗

∞X

=1

−1

+ ( + + + + ) (10)

where + = −1 Q

=0

+ , ≥ 1. The household’s perceived transversality condition,

7 **Staff** **Working** **Paper** **No**. **633** December 2016

lim

→∞ ∗ + −1 + =0

holds. This expression states that the present discounted value of consumption is equal

to the sum of perceived human **and** non-human wealth. 7

2.3 Firms

Our model economy features a continuum of large firms of measure ∈ [0 1]. Eachfirm

employs **labour** to produce consumption goods using the following technology

= (11)

where denotes output at the firm level, employment **and** ∈ (0 1] the aggregate

elasticity of output with respect to **labour**. In the baseline calibration, we assume constant

returns to scale, i.e. equal to 1, butlateronwealsoexaminethecaseofdecreasing

returns to scale in Section 6. The productivity innovation follows an exogenous process

given by

ln +1 = ln + +1 with ∼ (0) (12)

where ∈ (0 1) denotes the persistence of the technology process **and** is an i.i.d.

innovation with mean zero **and** st**and**ard deviation .

Sincepostingvacanciesiscostly,periodprofits, ,attime may be written as

= − − C ( ) (13)

where the number of job openings at the firm level **and** C (·) is a convex **and** increasing

vacancy cost function. The problem of each firm is to choose so as to maximise the

present discounted value of expected profits, which may be written as

max +

+

subject to the law of motion of employment at the firm level

∞X

=1

∗ −1

+ + for ≥ 0 (14)

+1 =(1− ) + (15)

The Bellman equation of this problem may be written as

J ( )=max

©

+ ∗

−1 J ( +1 ) ª

7 **No**te that profits are large under decreasing return but small under constant returns.

8 **Staff** **Working** **Paper** **No**. **633** December 2016

so that the problem of firm is to maximise the above equation subject to (11) **and** (15).

The first order condition with respect to is

C 0 ( )= ( ) ∗ −1

V +1 (16)

where V = J 0 ( ) denotes the marginal value of having an additional worker employed

at the firm. Equation (16) states that the marginal cost **and** benefit of posting a vacancy

must be equal. The envelope condition with respect to is

V = −1

− +(1− ) ∗ −1

V +1 (17)

This condition simply states that the value of having an additional worker employed at

the firm must be equal to the flow value - the marginal productivity of employment net

of wage costs - plus the continuation value of employment at the firm.

By leading equation (17) one period forward, multiplying both sides of the expression

by the stochastic discount factor, 1

, taking the expectation at time **and** combining the

resulting expression with equation (16), we obtain the following job creation condition

∙

¸

C 0 ( )

( ) = ∗

−1 +1 −1

+1 − +1 +(1− ) C0 ( +1 )

(18)

( +1 )

This condition is central to our analysis since it determines the optimal number of vacancies

that firm would like to post. The expression simply states that the expected

cost of a filled vacancy must be equal to its marginal benefit, which consists of expected

profits **and** savings generated from the additional match. This way of formulating the

firm’s problem means that the choice of current vacancies is based on the forecast of

future **labour** **market** conditions. The expected marginal cost of opening a vacancy can

be re-written in terms of the perceived sum of future profits per additional hire,

C 0 ∞

( )

( ) = X

(1 − ) −1 £

∗ +

−1 + −1

+ − +¤

(19)

=1

**No**te that this expression states that the value of an additional job must be equal to the

sum the future stream of profits that this job is expected to generate. In order to post

the optimal number of vacancies, firm must make forecasts up to the infinite future.

The firm perceives the following transversality condition holds:

∙

¸

lim (1 −

→∞ )−1

∗ +

−1 C 0 ( )

=0

( + )

9 **Staff** **Working** **Paper** **No**. **633** December 2016

2.4 Wage Negotiation

Wages are negotiated according to a Nash bargaining protocol. The wage maximises

the joint surplus of a match between workers **and** firms,

∙ H ( )

arg max

¸

(V ) 1−

where ∈ (0 1) denotes the workers’ bargaining power or, alternatively, the share of

surplus taken by the worker. 8

**No**te that the household’s surplus is the product of the

marginal value of having a member of the household employed at firm , defined in

equation (9), **and** (the inverse of the marginal utility of consumption). Dividing H()

by the marginal utility of consumption translates utility units of H in terms of goods.

The first order condition of this problem then yields the st**and**ard sharing rule that

characterises the optimal split of the aggregate surplus between workers **and** firms,

∙ H ( )

(1 − )

¸

= V (20)

To derive an expression for the bargained wage, , we assume that expression (20) holds

for subjective expectations (see equation (49) in Appendix A.1 **and** the details therein). 9

Furthermore, we assume that C ( ) takes a linear form, C ( )= ,asisst**and**ardin

the literature, where denotes the unitary vacancy cost. Thus, we obtain

= −1

+ +(1− ) (21)

The bargained wage is a weighted average of the marginal product of employment, the

cost of replacing the worker **and** the opportunity cost of working. 10

2.5 Aggregation **and** Market Clearing

We make the assumption that households **and** firms share the same set of beliefs about

the future. This assumption is reasonable because a) firms are owned by their employees,

b) household members **and** firms coordinate on expectations during the wage negotiation

process. This assumption of homogeneity in expectations across households **and** firms,

∗ = ∗ , implies a symmetric equilibrium i.e., = **and** = for all **and** .

8 The match values of the employment is specific to each member of the household **and** each employee

of the firm.

9 In line with what it is assumed under RE, this assumption makes it easy for the worker **and** the firm

to agree on the bargained wage. This is a convenient simplifying assumption for our purpose. There

may be alternative (more complex) ways of agreeing the bargained wage which we leave for future work.

10 As shown by Krause **and** Lubik (2007), the aggregate effects of intra-firm bargaining are negligible in

a st**and**ard search **and** matching framework with concave production functions. Thus, we abstract from

intra-firm bargaining in the analysis.

10 **Staff** **Working** **Paper** **No**. **633** December 2016

The **market** clearing condition in the goods **market** can be obtained by summing up

the period budget constraints **and** period profits (over )

+ = (22)

Imposing symmetry **and** following Mitra, Evans, **and** Honkapohja (2013), we assume

households use a consumption rule based on a linearisation of (10) around the steady

statevalues(¯ ¯ ¯ ¯ ¯). A tilde over a variable denotes the deviation of the variable

from its steady state value (i.e. ˜ = − ¯; where ¯ denotes the steady state value of ).

The linearised household’s behavior rule is given by

˜

1 − = ˜ (¯¯ +¯) 2

+¯ ˜ +¯˜ + ¯˜ +1 +˜ − ˜ +

1 −

"

#

−1

∞X

∞X

X

¯˜ + +

∗ ¯ ˜ + +˜ + − (¯¯ +¯) ˜ + (23)

=2

=1

We make the assumption that initial financial wealth is zero (**and** by symmetry it is

zero in all future periods). It is then convenient to re-write (23) as

where

˜

1 − =¯ ˜ +¯˜ + ¯˜ +1 +˜ −

=1

(¯¯ +¯) 2

˜ + S + S + S − S

1 −

(24)

S =

S =

∞X

∞X

¯ ∗ ˜ + S = ¯ ∗ ˜ +

=1

=2

"

#

−1

∞X

∞X X

∗ ˜ + **and** S =(¯¯ +¯)

∗ ˜ +

=1

=1 =1

The S S , etc variables denote the discounted sums of future forecasts **and** are key for

underst**and**ing the dynamic properties of the model that we explain at a later stage.

In line with the household problem, firms use a vacancy posting rule based on linearisation

of equation (19) around the steady state values of ¯, ¯, ¯, ¯ **and** ¯. The firms’

behavioral rule is therefore

∙ ¸

¯ 1¯˜ − ¯ + ¯ 1¯ ˜ 2 = ¯ 1 (1 − )˜ − (¯¯ 2 − ¯) 2 ∞X

1 − (1 − ) ˜ + (1 − ) −1 ∗¯ 1˜ + +

=2

"

∞X

(1 − ) −1

∗ ¯ 2˜ + − ˜ + − ¡¯¯ 2 − ¯ ¢ #

−1

X

˜ + (25)

=1

=1

11 **Staff** **Working** **Paper** **No**. **633** December 2016

where ¯ 1 = ( − 1)¯¯ −2 **and** ¯ 2 = ¯ −1 11 We re-write (25) as

µ

¯ 1¯˜ − ¯ + ¯ 1¯ ˜ 2 = ¯ 1 (1 − )˜ − (¯¯ 2 − ¯) 2

1 − (1 − ) ˜ + S + S − S − S (26)

where

S =

S =

∞X

∞X

(1 − ) −1 ¯2 ∗ ˜ + S = (1 − ) −1 ¯1 ∗ ˜ +

=1

=2

−1

∞X

∞X

X

(1 − ) −1 ∗ ˜ + **and** S =(¯¯ 2 − ¯) (1 − ) −1 +1 ∗ ˜ +

=1

=1

=1

The informational assumptions used in the derivation of equation (26) are detailed in Appendix

A.2. In particular, households **and** firms are assumed not to observe ˜ ˜ +1 ˜

**and** ˜ when they form their forecasts at time . The st**and**ard assumption in the **learning**

literature is to assume that agents’ forecasts at from period +1onwards are based

on past endogenous variables (here −1 −1 **and** −1 ). This approach conveniently

avoids the simultaneous determination of forecasts **and** endogenous variables. 12 This informational

assumption is particularly appealing here since the st**and**ard search literature

is calibrated to monthly frequency **and** the data is generally available with a lag; viewing

households **and** firms as having knowledge of contemporaneous monthly aggregate variables

when forming their forecasts is therefore implausible. Wage determination in the

search **and** matching literature is different from the st**and**ard neoclassical model **and** it

entails a more complex decision problem involving expectations. Under RE, households

**and** firms know that their expectations are the same. However, in our context, agents

may not know this to be true ex-ante. Ex-post, the bargained wage ˜ will depend on aggregate

variables such as ˜ ; see equations (21) or (27) below. In addition, the assumption

of time-to-hire is such that workers are matched to employeers in period but can only

start working in period +1. For all these reasons, the assumption of pre-determined

forecasts in this model is particularly plausible **and** we maintain this assumption in the

paper. An interpretation, owing to Evans **and** Honkapohja (2006), is that these forecasts

are obtained by households **and** firms from an econometric forecasting firm before going

to the **market** place. Armed with these forecasts, equations (24) **and** (26) provide a consumption

schedule for the representative household **and** a vacancy posting rule for firms

which they take to the **market** place. These (along with other equations outlined below)

determines a temporary equilibrium for the economy where contemporaneous monthly

11 **No**te that ¯ 1 is 0 **and** ¯ 2 is 1 when the production function exhibits constant returns to scale. For

brevity, we denote by b the absolute deviation of ( ).

12 The alternative approach would be to assume that current dated (here monthly) endogenous variables

are included in the forecasts of agents so that these endogenous variables **and** their forecasts are

determined simultaneously. As explained this approach is less realistic given the monthly frequency of

the model.

12 **Staff** **Working** **Paper** **No**. **633** December 2016

values of endogenous variables eg. ˜ ˜ ˜ ˜ ˜ **and** ˜ **and** ˜ +1 are all simultaneously

determined.

We turn to the other equations in the model. First, we linearise equation (21) around

the steady state **and** integrate it over to find an expression for the aggregate wage

˜ = ¯ 2˜ + ¯ 1˜ + ˜ +(1− ) ˜ (27)

Similarly, we linearise the unemployment rate (2), the production function (11), the

profit function(13),thelawofmotionofemployment(15)**and**thegoods**market**clearing

condition (22) to get the following conditions

˜ +˜ =0 (28)

˜ =¯ ˜ +¯¯ −1˜ (29)

˜ =¯¯ −1˜ +¯ ˜ − ¯ ˜ − ¯˜ − ˜ (30)

˜ +1 =(1− )˜ +¯˜ +¯˜ (31)

˜ + ˜ =˜ (32)

We also linearise equations (3), (4) **and** (12) to get

¯˜ +¯˜ =˜ (33)

˜ =( − 1) ¯ −2˜ (34)

˜ +1 = ˜ + +1 (35)

We now define a temporary equilibrium for this economy. A temporary equilibrium

is a set of values for the variables ˜ , ˜ , ˜ , ˜ , ˜ , ˜ , ˜ +1 , ˜ , ˜ , ˜ that, given the

exogenous stochastic process {˜ } ∞ = **and** the initial condition ˜ , satisfy the system of

equations consisting of (24) **and** (26)-(35). This equilibrium is determined as follows. In

every period given their forecasts, households enter the **market** with their consumption

schedule (24). Similarly, given their forecasts, firms enter the **market** with their vacancy

posting rule (26). These equations, together with the other equations outlined earlier

(which do not depend on forecasts), determine simultaneously the temporary equilibrium

values of ˜ , ˜ , ˜ , ˜ , ˜ , ˜ , ˜ +1 , ˜ , ˜ , ˜ at time . To complete the description of

temporary equilibrium one specifies how forecasts are formed. As argued before, in this

model it is plausible to assume that forecasts of firms **and** households are pre-determined

when they are brought to the **market**. The temporary equilibrium for the current period

provides a new data point for the agents. Given these new data, the forecasts of agents

are updated at the start of the following period using versions of recursive least squares

13 **Staff** **Working** **Paper** **No**. **633** December 2016

algorithm (which we explain later).

3 RE **and** Learning with Correctly Specified Beliefs

The RE solution of the model is of the following form

+1 = ¯ +¯ +¯ ˜ (36)

= ¯ +¯ +¯ ˜

= ¯ +¯ +¯ ˜

= ¯ +¯ +¯ ˜

since employment **and** productivity are the only state variables in this model. The RE

solution is expressed in levels rather than in deviations from their steady state values.

Under RE agents know that the productivity innovation follows an autoregressive process

**and** also the dispersion of that innovation. Rational agents have complete information

about the structure of the economy **and** know with certainty the true parameter values of

policy functions (denoted with a bar over the parameter). The solution of the RE model

is then used to forecast future values of the endogeneous variables. Therefore, rational

agents make no systematic mistakes. As is well known, the RE model is not able to

match the amplification present in the **labour** **market** data since productivity shocks are

unable to increase by much profitsperhireatthemargin. Moredetailedexplanations

are provided in Section 4.4.1.

Under adaptive **learning** (AL), agents are assumed to have perceived laws of motion

(PLMs) of the same form as equation (36) but do not have knowledge of the RE parameters

(¯ , ¯ **and** ¯ ), **learning** them as new information becomes available. In particular,

the correctly specified beliefs of agents consist of the following set of equations

= + + ˜ + for = { +1 } (37)

where are white noise processes. Agents are assumed to behave as econometricians **and**

estimate parameters **and** by running regressions of variables like employment,

profits, wages **and** interest rates on past employment **and** TFP innovations. This belief

system entails only a mild deviation from RE. Tables 8 **and** 9 in Appendix B show that

this belief system suffersfromthesameproblemastheREmodeli.e. itisnotable

to provide a solution to the unemployment volatility puzzle **and** unable to match the

forecast error properties in the data. 13 This seems to be the case since the correctly

specified beliefs converge to the RE solution. We, therefore, turn to a situation where

13 The computations in these tables use the same definitions used in Tables 3, 4 **and** 5; please refer to

sections 4 **and** 5.

14 **Staff** **Working** **Paper** **No**. **633** December 2016

agents use smaller forecasting models **and** consider if this can provide a resolution to

these problems.

Before doing so we first note that, following Campbell (1994), the RE solution (36)

may be written in an equivalent way. This RE solution involves as an AR(2) process

**and** as ARMA (2,1) processes. In particular, this solution can be shown to be

of the following form

+1 = 1¯ + 2 − 3 −1 +¯ (38)

= 1¯ + 2 −1 − 3 −2 +¯ +(¯ ¯ − ¯ ¯ ) −1

= 1¯ + 2 −1 − 3 −2 +¯ +(¯ ¯ − ¯ ¯ ) −1

= 1¯ + 2 −1 − 3 −2 +¯ +(¯ ¯ − ¯ ¯ ) −1

where 1 =(1− ¯)(1− ¯ ) 2 =(¯ +¯ ) **and** 3 =¯¯ .

Using correctly specified beliefs of the form (38) (or (36)) requires a lot of knowledge

from agents: they need to know that solutions for the endogenous variables are exactly

of the form above. In practice, ARMA type processes are significantly more difficult to

estimate. Recent experimental evidence by Hommes, Sonnemans, Tuinstra, **and** van de

Velden (2005) **and** Heemeijer, Hommes, Sonnemans, **and** Tuinstra (2009) suggests agents

estimate simple univariate autoregressive models to make forecasts about future variables.

Slobodyan **and** Wouters (2012) too show that univariate autoregressive forecasting

models can fit the data better than the RE model **and** are also closely related to the

survey evidence. Using AR type PLMs are hence very appealing. 14 For generality, we

also allow beliefs where agents estimate vector-autoregressive (VAR) models. This set

of beliefs is appealing since it allows interactions among the key endogenous variables

whilst being easy to estimate; it is also a popular forecasting tool in many

policy institutions. We deviate from RE in assuming that agents forecast the values of

the variables of interest based on simple AR or VAR belief specifications **and** examine

whether these belief specifications have the ability to get the search **and** matching model

closer to the data. Agents have incomplete knowledge about the structure of the economy

**and** they observe only their own objectives **and** constraints but do not observe other

agents’ preferences **and** beliefs. Thus, they do not know that their decisions are identical

to those of other agents.

14 Using AR (or VAR) beliefs obviates the need for agents to use productivity in their regression

equations (37). We remark that Slobodyan **and** Wouters (2012) use AR(2) PLMs as their preferred

specification in estimating medium-scale DSGE model based on small forecasting models. Their model

features a wider set of nominal **and** real frictions **and** the **dynamics** of their model is driven by multiple

innovations.

15 **Staff** **Working** **Paper** **No**. **633** December 2016

3.1 Learning with Autoregressive Beliefs

In this model we propose alternative beliefs systems of autoregressive form. 15

The simplest

forecasting model that is closest to the RE model is one where all endogenous

variables are forecasted using univariate autoregressive processes of order two i.e. AR(2)

processes. This serves as our benchmark for agents’ beliefs. In particular, we assume

that agents have Perceived Law of Motions (PLMs) of the form

= 0 + 1 −1 + 2 −2 + for = { +1 }

where are white noise processes. Like RE, the belief system is expressed in levels.

This belief specification represents only a modest departure from RE: the key difference

with the RE solution is that the moving average (MA) terms are dropped from

**and** (the employment equation is correctly specified). 16 This seems like a reasonable

assumption since wages, interest rates **and** profits are determined simultaneously by general

equilibrium considerations **and** depend on aggregate variables like ˜ : eg. firms do

not know the equilibrium mapping from these endogenous variables to **market** outcomes.

We now explain parameter updating under this set of beliefs. Let Φ =

³ 0 1 2´0

³

for = { +1 } **and** Ψ = 1 −1 −2´0.

We assume that agents use a

constant gain Recursive Least Squares (RLS) algorithm to update their beliefs

Φ = Φ −1 + Ψ −1

−1

³ −1 − Φ 0 −1Ψ −1´0

(39)

= −1 +

³Ψ −1 Ψ 0 −1 − −1´

where ∈ (0 1) denotes the constant gain **learning** parameter **and** is the precision (3 x

3) matrix associated with each equation. Constant-gain least squares is widely used in the

adaptive **learning** literature because it weighs recent data more heavily. See for example

Sargent (1999), Cho, Williams, **and** Sargent (2002), McGough (2006), Orphanides **and**

Williams (2007), Ellison **and** Yates (2007), Huang, Liu, **and** Zha (2009), Eusepi **and**

Preston (2011) **and** Milani (2011). Evans, Honkapohja, **and** Williams (2010) provide a

Bayesian justification for the constant gain RLS algorithm. In particular, they show that

constant gain RLS **learning** algorithm asymptotically approximates the Bayesian optimal

estimator when agents allow for drifting coefficients models. In addition, we assume that

15 These alternative belief specifications are similar in spirit to the simple wage rule models proposed

by Christiano, Eichenbaum, **and** Trab**and**t (2016) within the RE literature, where the wage rules are

autoregressive.

16 Under the proposed belief system the economy converges to a restricted perceptions equilibrium

(RPE) as in Sargent (1999), Cho, Williams, **and** Sargent (2002) **and** Branch **and** Evans (2006a) that is

different from the RE equilibrium (see Chapter 13.1 of Evans **and** Honkapohja (2001), Branch (2004) for

a discussion **and** Huang, Liu, **and** Zha (2009) for an application to the growth model). As our interest is

in matching unemployment volatility **and** survey forecasts, we do not study the nature of this RPE.

16 **Staff** **Working** **Paper** **No**. **633** December 2016

agents estimate the persistence parameter of the exogenous productivity process (12)

using constant gain RLS; this captures their uncertainty about the persistence of the

productivity process perhaps changing over time. 17

Agents update their beliefs over time by revising the value of parameters using a Recursive

Least Squares (RLS) algorithm. At the beginning of each period, agents inherit

the parameters of their belief system from the previous period, make forecasts **and** compute

the present discounted sums that allows them to form consumption **and** vacancy

posting decisions at every point in time. At the end of each period, agents are informed

about factor prices, (un)employment **and** profits **and** they update their beliefs in the following

period. In the **learning** literature, it is st**and**ard to assume that for the parameter

estimation agents use data available up to period − 1; eg. −1 −1 **and** −1 which

is what we do here.

We investigate an alternative set of beliefs that entails further deviations from RE. A

natural choice as a belief system is the univariate auto-regressive processes of order one,

i.e. AR(1) beliefs.InthiscasethePerceivedLawofMotions(PLMs)areoftheform

= 0 + 1 −1 + for = { +1 }

where are white noise processes. Agents use RLS algorithms in their estimation **and**

form their forecasts based on the AR(1) model using the informational assumptions stated

before.

A final set of beliefs we consider is one that allows for interactions between endogenous

variables. A typical way to allow for such interactions is to estimate vector auto-regressive

(VAR) models. Adopting multivariate autoregressive models to compute forecasts tends

to reduce the informational constraints that are present in simple univariate models. As

a result, agents can potentially capture (unknown) general equilibrium effects **and** reduce

forecast errors. In one sense, the AR specificationsendowagentswithmorespecific

knowledge about the form of the RE solution in that these PLMs do not allow for direct

dependence of any variable on the lags of other variables. The VAR specification, on

the other h**and**, allows agents to be relatively agnostic of the RE beliefs. Unlike the

univariate AR(2) **and**AR(1) belief specifications, the VAR(1) forecasting model allows

possible interactions among variables (employment, profits, wages **and** interest rates).

The VAR(1)neststheunivariateAR(1) model but not necessarily the AR(2)specification.

Since VARs have become st**and**ard forecasting tools in many macroeconomic institutions,

whilst being simple to estimate, we view this as as a plausible **and** attractive way for agents

17 In the AL literature, typically it is assumed that is known since it can be estimated consistently

from the exognenous process by RLS. Here we prefer to be realistic about the uncertainty firms may

face.

17 **Staff** **Working** **Paper** **No**. **633** December 2016

to form their forecasts in the **learning** set-up. For VAR(1), the PLM is of the form

= 0 + 1 −1 + for =( +1 ) 0

where 0 is a 4x1 matrix, 1 a 4x4 matrix **and** a vector of white noise processes.

Similarly, constant gain RLS algorithms are used to estimate the elements of matrices

0 **and** 1 in order for agents to form their forecasts. For economy of space, we do not

present the **learning** algorithms here.

4 Numerical Results

4.1 Calibration

We set the structural values of the parameters in the model following a st**and**ard calibration

exercise. First, we choose some parameter values using aprioriinformation. Second,

the choice of remaining parameters ensures that the stationary equilibrium of the model

matches a number of stylised facts as observed in the post-WWII US economy. As is

st**and**ard in the search **and** matching literature, a period in our model corresponds to a

month in the data.

The parameters chosen using aprioriinformation are the subjective discount factor

(), the exogenous separation rate (), the worker’s bargaining power (), **and** the elasticity

of the matching function with respect to vacancies (). The value of is set to

0996, which implies an annual real interest rate of about 4%. The value of is calibrated

to 0033 in order to match the evidence that jobs last on average two **and** a half years as

estimated in Davis, Haltiwanger, **and** Schuh (1996). We set the value of at 0.5 in line

with the literature. This value lies within the plausible interval of [05 07] as surveyed by

Petrongolo **and** Pissarides (2001). In order to facilitate comparability with the existing

literature, is chosen to be 05 (the choice of the values of **and** ensures that Hosios

(1990) condition is met.). The elasticity of output with respect to employment, , isset

to 1. Following Shimer (2010), we choose the persistence of the technology shock, , to

be 098 **and** the st**and**ard deviation of the productivity shocks to be 0005.

Theremainingthree**labour****market**parameters,namely, ¯ **and** , are set to match:

) a vacancy cost to output ratio of 0.01 in line with Andolfatto (1996), Gertler **and**

Trigari (2009) **and** Blanchard **and** Gali (2010);

) a vacancy filling rate of 27.8% as estimated by Shimer (2005), which is consistent

with a quarterly rate of 70% as in Trigari (2006) **and** den Haan, Ramey, **and** Watson

(2000);

18 **Staff** **Working** **Paper** **No**. **633** December 2016

Description Parameter Value

Discount Factor 0996

Parameter in the Utility Function 0882

Efficiency of the Matching Technology ¯ 0379

Elasticity of the Matching Function 05

Bargaining Power 05

Unitary Vacancy Posting Cost 0084

Separation Rate 0033

Elasticity of Output w.r.t. Employment 1

Productivity Level ¯ 1

Persistence of Productivity Shocks 098

St. Dev. of Productivity Shocks 0005

Gain Parameter 0002

Table 1: Calibrated Parameters - Monthly

) an unemployment rate of 6%, which corresponds with the st**and**ard ILO definition

of unemployment for the post-WWII US average.

The resulting replacement ratio — computed as the disutility of work over the wage

**and** the marginal utility of consumption () — is equal to 83%, which is above the

value of 70% suggested by Mortensen **and** Nagypal (2007). According to a study by

Hagedorn **and** Manovskii (2008), a total replacement ratio of around 95% can generate

**labour** **market** fluctuations that are in line with the empirical evidence. Their study argues

that, if the outside option of the worker is high (this happens when both is low **and**

the replacement ratio high), then (steady state) firm’s profits are small **and** can generate

greater amplification in **labour** **market** variables. The replacement ratio in their study

is equal to the value of non-**market** activity that includes both unemployment subsidies

**and** home production. Costain **and** Reiter (2008) have criticised this calibration because

it gives rise to excessive sensitivity of unemployment to unemployment benefits. The

resulting replacement ratio ensures that our results are not driven by the Hagedorn **and**

Manovskii effect. This will become clear once we examine the amplification properties of

the model under RE beliefs. Table 1 provides a summary of the parameters used in the

baseline calibration of our hypothetical model economy.

We choose the gain parameter in the **learning** algorithm to be =0002 (equivalent

toavalueof0006 in the corresponding quarterly model), which implies that agents use

past data to update their beliefs for around 42 years (10002 = 500 months). There

is lack of consensus in the **learning** literature concerning the constant gain parameter,

which ranges from 0002 to 0035 at quarterly frequencies. See for example Eusepi **and**

Preston (2011), Branch **and** Evans (2006b), Milani (2007) **and** Orphanides **and** Williams

(2007). The value chosen for this parameter is relatively small because we exclude policy

considerations (for example as in Mitra, Evans, **and** Honkapohja (2016) or Mitra, Evans,

**and** Honkapohja (2013) where a higher gain parameter is used) but lies within the range

19 **Staff** **Working** **Paper** **No**. **633** December 2016

ˆ ˆ ˆ ˆ

ˆ ˆ

1 ˆ

1 057 940 884 1782 079

(b 1 b 2−1 ) 081 090 091 089 091 080

ˆ 1 079 083 −078 083 085

ˆ − 1 090 −097 095 063

(b 1 b 2 ) ˆ − − 1 −091 098 069

ˆ − − − 1 −098 −061

ˆ − − − − 1 066

ˆ − − − − − 1

Table 2: Summary Statistics, Quarterly US Data

**No**tes. Relative st**and**ard deviations, autocorrelation **and** correlation coefficients in this Table correspond to the quarterly

data series detrended using a Hodrick-Prescott filter with smoothing parameter 1600. Each data series x correspond to a

variable in the model. The term ( 1 2 ) st**and**s for the correlation coefficient between variables 1 **and** 2

of parameters suggested in the literature. We conduct a robustness exercise with different

values of the gain parameter in Section 6.

4.2 US Data

In this section we compare the main statistical properties of the simulated **labour** **market**

series generated from the model with the corresponding series in the US data, in particular

focusing on second moments. It is st**and**ard practice, in one-worker-one-firm models, to

compute the st**and**ard deviations of **labour** **market** data relative to productivity, but, in

large-firm models, the st**and**ard measures that capture the volatilities of data are typically

divided by the corresponding measure of output. Since our model is a large-firm model,

we take the latter approach to compute relative st**and**ard deviations.

The seasonally adjusted series of (un)employment is taken from the Bureau of Labour

Statistics (BLS). As a proxy for vacancies we merge the seasonally adjusted help-wanted

advertising index released by the Conference Board with the vacancy series calculated by

Barnichon (2010). Aggregate output is measured as seasonally adjusted real GDP, which

is drawn from the National Income **and** Product Account (NIPA) tables 116 **and** 115.

All data series are quarterly **and** cover the period ranging from 19511 to 20144 (due

to the data availability on vacancies). Table 2 summarises the main cyclical properties

of the logged detrended series.

One of the most salient features in the data is the high volatility of unemployment,

vacancies **and** **labour** **market** tightness as reported in Table 2. In particular, both vacancies

**and** unemployment are about 940 **and** 884 times more volatile than the aggregate

output respectively. Moreover, the measure of **labour** **market** tightness is around 1782

timesmorevolatilethanoutput. 18

18 Another well known stylised fact in **labour** **market** **dynamics** is the negative relationship between

vacancies **and** unemployment, also known as the Beveridge curve.

20 **Staff** **Working** **Paper** **No**. **633** December 2016

4.3 Simulation Results

We simulate a search **and** matching model under the different belief specifications **and**

compare results. We use st**and**ard methods to solve **and** simulate the RE model. We

initiate the simulations of our **learning** models from values consistent with the deteministic

steady state **and** then generate a series for 10 900 periods using the **learning** algorithm

previously stated. The first 10 000 periods ensure convergence **and** are discarded. We

keep the remaining 900 observations, which correspond to 75 years of data, so as to

guaranteethatthesimulatedseriesarefreefrom any transitional dynamic considerations.

We then repeat this procedure 1 000 times **and** report the mean values of the variables

of interest. 19 Since our model is calibrated for monthly frequencies **and** the US data is

reported only in quarterly frequencies, we then convert the monthly simulated series into

quarterly frequencies. Finally, we transform the simulated series from deviations into

percentage deviations by dividing the simulated series by the corresponding steady state

values.

Tables 3 **and** 4 report the statistical properties of the simulated series of interest

under **learning** for all three belief specifications. In the tables **and** figures below, a hat

over a variable denotes the percentage deviation of the variable from its steady state

value (eg. ˆ =˜¯). Table 3 shows that the RE model generates very little amplification

in **labour** **market** variables. The table shows that the search **and** matching model with

**learning** can replicate the second moments of the US **labour** **market** remarkably well.

All the **learning** models provide a good match for the relative volatility in vacancies,

unemployment **and** labor **market** tightness. For instance, in the AR(2) specification,

vacancies **and** unemployment are about 887 **and** 700 times more volatile than output

respectively with the corresponding numbers in US data being 930 **and** 884 All the

**learning** models do significantly better than the RE model in matching amplification in

the data. The correlations **and** autocorrelations in the VAR(1) model are, however, worse

than those in the AR(1) **and**AR(2) specifications.

4.4 Impulse Responses

In this sub-section, we study how **labour** **market** variables respond to a TFP innovation

**and** compare the **dynamics** under RE **and** **learning**. Following Eusepi **and** Preston (2011),

the impulse responses of the **learning** model are computed by simulating the model twice

over 10 000+120 periods. We add to the first simulation a positive 1% st**and**ard deviation

19 We check the stability of each **learning** model at each point in time by examining the highest eigenvalue

of the system **and** when this stability condition is not met we disregard the entire draw. We discard

around 12% of the draws in the univariate AR models **and** around 23% in the VAR(1) model. The initial

values of parameter estimates in each regression equation is typically set at zero with the intercept terms

set equal to the corresponding deterministic steady value of the variable. The precision matrices are set

equal to the identity matrix to start the simulations. We have experimented with other initial values

**and** our results do not change significantly.

21 **Staff** **Working** **Paper** **No**. **633** December 2016

Model Statistics ˆ ˆ ˆ ˆ

ˆ ˆ

ˆ1 ˆ 1 003 050 043 091 100

(b 1 b 1−1 ) 096 097 092 097 096 096

ˆ 1 099 099 −099 100 100

RE ˆ − 1 096 −100 099 099

(b 1 b 2 ) ˆ − − 1 −096 099 099

ˆ − − − 1 −099 −099

ˆ − − − − 1 100

ˆ − − − − − 1

ˆ1 ˆ 1 058 1140 916 1985 092

(b 1 b 1−1 ) 095 094 071 094 089 094

ˆ 1 095 083 −096 092 100

AR(1) ˆ − 1 094 −099 099 093

(b 1 b 2 ) ˆ − − 1 −086 097 079

ˆ − − − 1 −096 −095

ˆ − − − − 1 089

ˆ − − − − − 1

Table 3: Summary Statistics: RE **and** AR(1) **learning** model

**No**tes. Relative st**and**ard deviations, autocorrelation **and** correlation coefficients in this Table correspond to the quaterly

simulated series expressed in percentage deviations from steady state value. The term ( 1 2 ) st**and**s for the correlation

coefficient between variables 1 **and** 2

Model Statistics ˆ ˆ ˆ ˆ

ˆ ˆ

1 ˆ

1 044 887 700 1521 094

(b 1 b 1−1 ) 095 094 070 094 089 094

ˆ 1 094 080 −095 090 100

AR(2) ˆ − 1 092 −098 099 092

(b 1 b 2 ) ˆ − − 1 −083 096 076

ˆ − − − 1 −095 −094

ˆ − − − − 1 088

ˆ − − − − − 1

1 ˆ

1 031 867 483 1160 099

(b 1 b 1−1 ) 089 059 −007 061 035 088

ˆ 1 064 030 −071 052 100

VAR(1) ˆ − 1 077 −091 095 059

(b 1 b 2 ) ˆ − − 1 −043 093 022

ˆ − − − 1 −074 −069

ˆ − − − − 1 045

ˆ − − − − − 1

Table 4: Summary Statistics: AR(2) **and** VAR(1) **learning** models

**No**tes. Relative st**and**ard deviations, autocorrelation **and** correlation coefficients in this Table correspond to the quaterly

simulated series expressed in percentage deviations from steady state value. The term ( 1 2 ) st**and**s for the correlation

coefficient between variables 1 **and** 2

22 **Staff** **Working** **Paper** **No**. **633** December 2016

(productivity) shock in period 10 001 **and** compute the impulse responses as the difference

between the two resulting set of impulse responses from period 10 001 onwards. This

experiment is then repeated 1 000 times **and** the mean impulse responses of the variables

of interest are reported. The simulated series are converted into quarterly frequencies

**and** then expressed in percentage deviations from steady state values.

Figure 2 shows the impulse responses of aggregate output, (un)employment, vacancies,

**labour** **market** tightness **and** the job filling rate to a positive TFP innovation under

**learning** **and** RE. The impulse responses under RE display negligible amplification relative

to the **learning** models. Firms endowed with RE beliefs correctly underst**and** the general

equilibrium restrictions that determine future wages. For this reason, expected wages tend

to absorb most of the productivity increase **and**, as a result, unemployment, vacancies

**and** the measure of **labour** **market** tightness respond only marginally to a TFP innovation.

Very distinct dynamic responses are observed under **learning** to a positive TFP innovation

for all three belief specifications. Following a positive technology shock, the

incentive for vacancy creation increases sharply on impact, leading to more employment

**and** a sharp fall in unemployment. A productivity innovation thus leads to more vacancy

creation under AL. The response of employment is largest for the AR(1) beliefs followed

by AR(2) **and**VAR(1) beliefs; the amplification generated in the **learning** models is, however,

much larger than that of the RE specification. Quantitatively, the magnitude of

the responses is preserved for the next four years, i.e. it is greater under AR(1) than

under AR(2) **and**VAR(1). After four years, the magnitude of these responses is reversed

i.e. it is smaller under AR(1) than under AR(2) **and**VAR(1). We observe over **and**

under-shooting of the variables under AL compared to RE as they converge to the steady

state.

23 **Staff** **Working** **Paper** **No**. **633** December 2016

0.03

ŷ

0.02

̂n

0.6

̂v

0.02

0.01

0

0.01

0

0.4

0.2

0

−0.01

−0.01

0 20 40 0 20 40

0.1

û

1

̂θ

−0.2

0 20 40

0.4

̂q

0

−0.1

−0.2

0.5

0

0.2

0

−0.2

−0.3

0 20 40

−0.5

0 20 40

RE AR(1) AR(2) VAR(1)

−0.4

0 20 40

Figure 2: Impulse responses to **labour** **market** variables. **No**tes. Impulse responses to a 1% st**and**ard deviation

(productivity) shock. The solid **and** dotted lines show the impulse responses for the key **labour** variables under RE **and** AL.

Percentage deviations from steady state values reported along the vertical axis. The horizontal axis display the number of

quarters after the shock.

Figure 2 shows that **learning** increases the internal propagation mechanism of the

model. In particular, the response of output significantly exceeds that of the TFP innovation

on impact. In contrast, under RE the response of output is approximately the

same as the TFP innovation, which is a feature of the unemployment volatility puzzle.

The reason behind the greater response in output under **learning** has to do with a larger

increase in employment. The **dynamics** of consumption (not plotted in the figure) is

qualitatively the same as that of output.

Intuitively, as the **labour** **market** tightens **and** the job finding probability falls, both

the marginal costs **and** the benefits associated with filling an additional vacancy tend

to increase. The response in unemployment is, however, more sluggish relative to the

response in vacancies given that unemployment is a pre-determined variable.

24 **Staff** **Working** **Paper** **No**. **633** December 2016

Ŝ f

0 20 40

0.5

1

Ŝ z f

0

0.5

−0.5

0 20 40

0

1

Ŝ w f

0.01

Ŝ r f

0.5

0

−0.01

0

0 20 40

−0.02

0 20 40

RE AR(1) AR(2) VAR(1)

Figure 3: Impulse responses to infinite sums in job creation condition. **No**tes. Impulse responses to a 1%

st**and**ard deviation (productivity) shock. The solid **and** dotted lines show the impulse responses for the infinite sums in the

job creation condition under RE **and** AL. Percentage deviations from steady state values reported along the vertical axis.

The horizontal axis display the number of quarters after the shock.

To disentangle further these effects under **learning** **and** gain intuition, we rearrange

the job creation condition, equation (26), in linearised form. For simplicity, we set

equal to 1, inwhichcase¯ 1 =0**and** ¯ 2 =1**and**, hence, S =0.20 Thus, we have

− ¯ ˜ 2 = S ≡ S − S − S (40)

**and**weplottheimpulseresponsesoftheinfinite sums in (40). As shown earlier, the

infinite sum S , which denotes the present discounted value of profits per hire, responds

much more strongly under AL relative to RE. Figure 3 decomposes this infinite sum, S ,

into the three components, S , S **and** S . **No**te that S is the same under RE **and**

**learning** **and**, hence, the lines overlap. This figure shows that the order of magnitude of S

is much smaller relative to that of S **and** S (note the vertical access of S ); the sum of

stochastic discount factors displays little variation relative to the other two infinite sums. 21

**No**te that the size of the responses of both S **and** S under RE are of equal magnitudes,

which explains the negligible amplification generated by the st**and**ard RE search **and**

matching model. In sharp contrast, the magnitude of the response of S is much higher

20 **No**te that wage bargaining turns out to be the same for small **and** large firms when =1.

21 This result is suggestive that a model where households’ preferences are linear in consumption is

likely to generate similar results.

25 **Staff** **Working** **Paper** **No**. **633** December 2016

than that of S after a TFP innovation under **learning** for all belief specifications. The

difference between S **and** S explains the strong amplification mechanism of the **learning**

models **and** the high variation in vacancy posting on impact. The extent of the difference

in these two magnitudes, **and** hence amplification, depends on the forecasting models

used by agents.

4.4.1 Mechanism of Amplification

We now examine the transmission mechanism of TFP innovations under different belief

specifications. We first describe the mechanism under RE beliefs in order to underst**and**

why the model fails to generate amplification in **labour** **market** variables **and** then provide

an intuition as to why the **learning** models are able to provide a solution for the

unemployment volatility puzzle.

A TFP shock has well-understood implications in the st**and**ard search **and** matching

model under RE beliefs. A positive innovation, +1 , shifts the production frontier **and** the

**labour** dem**and** schedule, by increasing **labour** productivity, which raises marginal profits

per hire. Since wages in the st**and**ard model are flexible **and** negotiated through the

process of Nash bargaining, a shift in technology increases both future marginal products

of employment **and** wage costs leading to more vacancy creation. However, the extent to

which the model is able to generate amplification depends on the way in which wages are

determined.

Nash bargained wages tend to absorb most of the increase in productivity under RE,

leading to little incentive for vacancy creation. To get further intution, we start from

equation (40). We ignore the term involving S in equation (40), since it plays a minor

role, **and** re-write equation (40) approximately as

− ˜

¯ ≈ ¯

∞X

(1 − ) −1 [˜ + − ˜ + ] ≡ ¯

=1

¡ ¢

S

− S

(41)

We replace

∗ by , the expectation operator under RE. Since unemployment is a

pre-determined variable, this expression pins down how many vacancies are posted in

equilibrium. **No**te that vacancy creation depends on the present discounted value of

profits per hire, which responds only little under RE. The RE solution expressed in terms

of deviations from steady state values may be written as

˜ +1 = ¯ ˜ +¯ ˜

˜ = ¯ ˜ +¯ ˜

˜ +1 = ˜ + +1

26 **Staff** **Working** **Paper** **No**. **633** December 2016

where ¯ 1 2

denotes the elasticities of variable 1 with respect to variable 2 . For the

baseline calibration, 0 ¯ 1 2

1 for all possible combinations of 1 **and** 2 . Plugging

the RE solution into (41) yields

where

¯ 1 =

− ˜

¯ ≈ ¯ ¡¯1˜ −

¯ 2˜ − ¯

¢

3˜ (42)

¯

1 − (1 − )¯ , ¯2 =

¯ ¯

1 − (1 − )¯

**and** (43)

¯ 3 = ¯ ¯ (1 − (1 − ) )+¯ (1 − (1 − )¯ )

(44)

[1 − (1 − )¯ ][1− (1 − ) ]

This expression provides the key insight into the lack of amplification under RE.

At the outset, note that the vacancy posting decision does not respond to employment

because ˜ is pre-determined; it is next period employment that affects vacancy posting.

There is a direct effect from productivity innovations, ˜ , to the vacancy posting decision

through shifts in future returns from employment (term ¯ 1˜ in equation (42), which

corresponds to the first infinite sum in (41)) **and** an indirect effect from productivity

innovations through shifts in future wage costs (the terms ¯ 2˜ **and** ¯ 3˜ in equation (42),

which correspond to the second infinite sum in (41)). For the baseline calibration, it can

be shown that ¯ 1 **and** ¯ 3 in equation (42) are very close to each other, which implies that

productivity innovations have little impact on the present discounted value of profits per

hire. 22 In other words, when firms correctly forecast future profits per hire, the variability

of profits per hire is dampened because wages are flexible **and** absorb large part of the

innovation. As a result, the RE model is unable to generate sufficient amplification in

vacancies.

We now turn to the explanation of amplification under **learning**. We first note that

under RE firms have knowledge of the general equilibrium restrictions that determine

future wages **and** employment; in particular, that a positive productivity shock leads

to higher wages **and** employment in the future. Under **learning**, on the other h**and**,

firms are not aware of these general equilibrium restrictions, i.e. that future wages **and**

employment will be higher due to higher productivity. In the **learning** model agents are

unable to underst**and** the implications of future wage negotiations **and** instead estimate

simple autoregressive models to forecast future wages using past information.

To keep the explanation simple, we restrict ourselves to a discussion of the results

under **learning** with AR(1) beliefs. Similar logic applies to the AR(2) **and**VAR(1) belief

specifications. Under AR(1) beliefs,firms estimate the parameter 1 by running the

following regression ˜ = 1 ˜ −1 , to make their forecasts of future wages. Combining

22 The baseline calibration ensures that our results are not driven by the HM strategy; see Hagedorn

**and** Manovskii (2008).

27 **Staff** **Working** **Paper** **No**. **633** December 2016

this expression with (41) yields

− ˜

¯ ≈ ¯

( 1˜ − 2 ˜ −1 ) (45)

where 1 = **and**

1−(1−) 2 =

2 1

1−(1−) 1

. A TFP innovation increases the discounted

value of future marginal products of employment, the term 1˜ in equation (45), but has

no direct impact effect on the present discounted value of future wage costs, the term

2 ˜ −1 . 23 The high persistence of the TFP innovation is key for generating amplification

under**learning**becauseitraisestherateatwhichfuturemarginalproductsofemployment

are discounted. This greater persistence together with the fact that expectations are predetermined

implies that firms overestimate the present discounted value of profits per

hire, which then translates into further incentives for vacancy creation. Thus, the **labour**

**market** becomes tighter **and** the job filling rate falls after a TFP innovation, which then

increases the expected costs of vacancy posting. After a technology innovation, expected

marginal products of employment increase more than expected wage payments. More

vacancies in equilibrium translate into higher employment **and** output. Figure 3 shows

that the discounted value of future costs (wages) is lower than the discounted value of

future marginal products under AL.

To provide further clarity, Figure 4 plots the path of simulated S − S (i.e. the

present discounted value of profits per hire in the model with constant returns to scale)

for the different models over 300 quarters. Under RE the two sums are very close to

one another, explaining the lack of amplification in the model (note the magnitude on

the vertical axis). The figure shows that the difference in the two sums is greatest under

AR(1) beliefs followed by AR(2) **and**VAR(1) beliefs. **No**te, however, that the order

of magnitude in the discounted value of profits per hire even under VAR(1) beliefs is

significantly higher than under RE beliefs. This means that **learning** models can help

better match the relative volatility of **labour** **market** variables with the US data. The

persistence of the simulated series is highest under AR(1) **and**AR(2) beliefs, helping

explain the greater propagation in vacancy creation.

23 Recall that the elasticity ¯ 1 under RE differs from the elasticity 1 under **learning** because is

estimated at each point in time.

28 **Staff** **Working** **Paper** **No**. **633** December 2016

2

x 10 −4

RE

0.05

AR(1)

0

−2

0

0.02

0

−0.02

50 100 150 200 250 300

AR(2)

50 100 150 200 250 300

−0.05

Ŝ z f -Ŝw f

0.02

0.01

0

−0.01

−0.02

50 100 150 200 250 300

VAR(1)

50 100 150 200 250 300

Figure 4: Simulatedpathofinfintesumsinthejobcreationcondition. **No**tes. Infinite sums are expressed

in percentage deviations from state state values. The first 10000 months are discarded **and** the remaining 900 months

averaged over 500 replications under RE **and** AL. Monthly series are converted into quarterly series.

5 Forecast Properties of the Model

Vacancy posting decisions in the search **and** matching models typically depend on the

marginal profitability of long-term employment relationships. Thus, (un)employment **and**

wage forecasts play a key role for job creation. Under RE firms can perfectly assign a

value to a filled vacancy because they have full information **and** knowledge about the

structure of the economy. As shown in Section 3, relaxing this assumption can generate

greater amplification in the value of a filled vacancy. Given that forecasts are central for

explaining amplification in **labour** **market** data, a suitable test of the AL models would

be to compare how well the statistical properties of forecasts generated by the AL models

(**and** RE) compare with those in the data. As mentioned in Section 1, adaptive **learning**

models have been used to successfully match macro **and** survey data (see e.g. Adam,

Marcet, **and** Beutel (2016), Kuang **and** Mitra (2016), Slobodyan **and** Wouters (2012),

Ormeno **and** Molnar (2015) **and** Eusepi **and** Preston (2011)). Since wage forecasts are not

available in the Survey of Professional Forecasters (SPF), we focus only on unemployment

forecasts. We show that our **learning** models are more successful than RE at matching

survey data on unemployment forecasts.

As indicated by the PLMs, agents forecast all future variables using simple autoregres-

29 **Staff** **Working** **Paper** **No**. **633** December 2016

sive models. Since agents in our model are required to forecast (un)employment rates, we

can therefore compare the performance of forecast errors of unemployment rates relative

to the data. We take the quarterly forecasts of unemployment rates from the SPF, which

are collected by the Federal Reserve Bank of Philadelphia. Questionnaires are sent to the

individual forecasters to carry out their expert forecasts at the end of the first month of

each quarter **and** individual forecasts are released towards the end of the second month

of each quarter. Forecast data are available from 19684 to 20144 **and**bothmean**and**

median of forecasts are reported up to four quarters ahead. We compute the forecast

errors by removing the log of the actual realisation from the log of the forecast data of

the unemployment rates up to four periods ahead using BLS data.

Table 5 reports the statistical properties of forecast errors, including their relative

volatility with respect to detrended output, autocorrelation of the forecast errors **and**

their correlation with the first differences of output **and** unemployment. We find that the

forecast errors are much more volatile relative to output over the 4-quarter horizon. The

table shows that the one quarter-ahead forecast errors is almost 456 times more volatile

than output. In addition, the quarterly forecast errors display high serial correlation,

which is stronger as the time horizon of the forecast increases. Furthermore, over the

business cycle, the forecast errors of unemployment are all pro-cyclical, which means

agents tend to over-predict unemployment during expansions **and** under-predict it in

recessions.

The statistical properties of the simulated forecast errors, expressed in percentage

deviations from the steady state unemployment rate under both RE **and** AL, are also

reported in Table 5. The simulation methods used here are the same as described in

Section 4.4. The quarterly **and** yearly forecast errors of unemployment are computed as

the three **and** twelve months ahead forecasts. Consistent with SPF the quarterly **and**

yearly forecasts are conducted in the second month of the quatter. It follows from the

table that the performance of the RE model at matching the statistical properties of the

forecast errors in the data is rather poor. The RE model predicts a very low relative

volatility of the forecast errors, negligible autocorrelation **and** no pro-cyclicality.

The **learning** models perform better than the RE model in matching the statistical

properties of forecast errors. In particular, the relative volatility of forecast errors vis-avis

output are better matched by all the **learning** models; in sharp contrast RE does a

poor job in this regard. AR(1) **and**AR(2) modelsmatchthefirst order autocorrelations

of forecast errors two to four quarters ahead than the RE **and** VAR(1) models. The

models have less success in matching the first order autocorrelation of the one-step ahead

forecast errors; the AR(2) model is the best in this regard. The correlations between the

forecast errors **and** the changes in unemployment **and** output are less well matched with

the data; qualitatively, the AR(2) matches the signs of the correlations. 24 Arguably it

24 In an early version Di Pace, Mitra, **and** Zhang (2014), we show that the forecast error properties in

30 **Staff** **Working** **Paper** **No**. **633** December 2016

1

2

3

4

ˆ1 ˆ 456 621 802 989

Data (ˆ 1 ˆ 1−1 ) 070 079 085 089

(ˆ 1 ∆ˆ ) −069 −055 −047 −036

(ˆ 1 ∆ˆ ) 044 042 037 032

ˆ1 ˆ 005 012 018 021

RE (ˆ 1 ˆ 1−1 ) 013 024 061 073

(ˆ 1 ∆ˆ ) −010 −004 −003 000

(ˆ 1 ∆ˆ ) 027 000 004 00

ˆ1 ˆ 338 511 650 742

AR(1) (ˆ 1 ˆ 1−1 ) 007 073 084 090

(ˆ 1 ∆ˆ ) 000 −014 −019 −018

(ˆ 1 ∆ˆ ) 012 020 024 023

ˆ1 ˆ 282 397 479 542

AR(2) (ˆ 1 ˆ 1−1 ) 033 070 084 088

(ˆ 1 ∆ˆ ) −010 −011 −015 −017

(ˆ 1 ∆ˆ ) 014 014 020 021

ˆ1 ˆ 393 449 471 480

VAR(1) (ˆ 1 ˆ 1−1 ) 013 053 059 061

(ˆ 1 ∆ˆ ) 008 −002 003 −012

(ˆ 1 ∆ˆ ) 002 010 013 020

Table 5: Forecast Properties

**No**tes. The term ( 1 2 ) st**and**s for the correlation coefficient between variables 1 **and** 2 . Data from Survey of

Professional Forecasters. Forecast errors are defined as

=ln +

− ln ( + ) for =1 2 3 **and** 4

data, where u F t+j

denotes forecasted unemployment,.**and** as

= ˆ + − ˆ + for =1 2 3 4 in the models.

is remarkable that our simple models with **learning** are able to match the data along so

many dimensions especially in terms of amplification **and** forecast error properties.

5.1 Wage Rigidity under RE as an alternative?

As mentioned in Section 1, the introduction of wage rigidity in the st**and**ard search **and**

matching model under RE is able to generate greater amplification in **labour** **market**

variables. The rationale behind this result is simple: if quantities display little amplification

after TFP innovations, fixing prices have the potential to generate the expected

results. We now show that, while the introduction of wage rigidities leads to higher

volatility under RE (as is well known in the literature), the properties of forecast errors

of unemployment are less well matched to those of the SPF.

In accordance with Hall (2005) a wage norm or social consensus can be perceived as

a rule to select an equilibrium for the wage within the bargaining set. We assume that

the wage is given by a weighted average of the steady state wage ( ¯) **and**thenegotiated

unemployment are better matched by an alternative set of beliefs where agents regress the key endogenous

variables only on unemployment (but not on the productivity shock). This is perhaps due to the fact

that agents only use unemployment rates in this alternative version unlike our current **learning** models.

31 **Staff** **Working** **Paper** **No**. **633** December 2016

Relative Volatilities

ˆ ˆ ˆ

ˆ ˆ

1 ˆ 038 697 600 1284 094

Forecast Error Properties

Statistics 1

2

3

4

ˆ1 ˆ 058 173 248 305

(ˆ 1 ˆ 1−1 ) −004 025 062 076

(ˆ 1 ∆ˆ ) 000 −007 −008 −009

(ˆ 1 ∆ˆ ) −012 003 007 008

Table 6: Wage Rigidities **and** RE

**No**tes. Relative st**and**ard deviations this table correspond to the quarterly simulated series expressed in percentage

deviations from the steady state values in the RE model featuring wage rigidities. Forecast errors

= {1 2 3 4} as defined in Table 5.

for

wage under the Nash protocol in equation (27),

˜ = ¯ +(1− ) ˜

where ˜

is the bargained wage in (27) **and** ∈ (0 1) denotes the wage rigidity

parameter. We set the value of to 07 in line with Blanchard **and** Gali (2010).

Table 6 reports the relevant statistics for this model. Amplification in **labour** **market**

variables **and** the forecast error properties are much improved relative to the RE model

with flexible wages. Nevertheless, **learning** models continue to perform better in terms

of amplification in **labour** **market** variables (which are closer to the data) relative to

the RE model with wage rigidity. 25 In addition, they provide a much better fit tothe

forecast error properties in the data than RE. For this reasons, the **learning** models are

our preferred relative to sticky wages under RE. In one sense, the **learning** models maybe

thought of a way to rationalise wage rigidity under RE with the distinction that, while

under **learning** wages respond fully to TFP innovations, wage expectations are sluggish.

6 Robustness

In this subsection we look at the sensitivity of our results to alternative parameterisations

of the **learning** models. We first consider a calibration of the model with decreasing

returns to scale in **labour** **and** then consider two values of the constant gain parameter.

For economy of space, we choose to focus on the AR(2) specification for our robustness

exercise.

We consider two alternative values of the gain parameter to check the sensitivity

25 We have experimented with higher values of **and** we find that relative volatilities of **labour** **market**

variables are in line with the data but the correlations remain the same. In addition, wage norms that

depend on past wages, rather than on steady state values, the RE model struggles to match the data

along all dimensions.

32 **Staff** **Working** **Paper** **No**. **633** December 2016

ˆ ˆ ˆ ˆ ˆ ˆ ˆ

ˆ ˆ ˆ

Data 057 930 884 1771 079

Alternative parameterisations

AR(2) (baseline) 044 887 700 1521 094

AR(2) ( =0001) 044 894 695 1519 094

AR(2) ( =0003) 044 975 687 1544 095

AR(2) ( =07) 043 978 681 1525 094

AR(2) ( =06 **and** =04) 050 875 776 1457 096

Table 7: Alternative Parameterisations

**No**tes. Relative st**and**ard deviations in this Table correspond to the quarterly simulated series expressed in percentage

deviations from steady state values under different parameterisations.

of our results to the way in which agents discount past information. We choose the

following two parameters of : 0001 **and** 0003. Table 7 reports the relative st**and**ard

derivations of **labour** **market** variables under the two alternative parameterisations (along

with the results for the baseline AR(2) model). Our simulations show that the values

of the gain parameter have very little influence in terms of the statistical properties of

the simulated data. A smaller gain parameter indicates that agents put more weight on

past data relative to current data to update beliefs, which means that the time it takes

for the effects of TFP innovations to vanish from agents’ information set is longer. The

amplification results are therefore not much affected by the choice of the parameter .

The calibration of our baseline model assumes constant returns to scale in **labour**.

In line with models featuring physical capital, we set the value of equal to 07 **and**

simulate the model under this new parameterisation **and** check the sensitivity of the

robustness of the results. The main conceptual difference with the baseline calibration

is that the dem**and** curve under decreasing returns to **labour** is downward sloping rather

than perfectly elastic. We find that the results from this simulation are in line with the

simulations of our baseline AR(2) model. Table 7 shows that the **labour** **market** variables

still display a great deal of variation.

Finally we carry a robustness exercise with regards to the elasticity of the matching

function with respect to vacancies () **and**thebargainingpowerofworkers(). In carrying

out this exercise we ensure that Hosios condition is met by increasing the elasticity

of the matching function with respect to vacancy **and** reducing at the same time the bargaining

power of workers by 01 relative to the baseline calibration. Table 7 shows that

this alternative calibration delivers robust results under AR(2) beliefs. Table 10 in the

Appendix shows that the forecast errors properties under **learning** are unchanged under

these robustness exercises i.e. the **learning** models continue to provide a good match to

the data.

33 **Staff** **Working** **Paper** **No**. **633** December 2016

7 Conclusion

In the st**and**ard search **and** matching model the vacancy posting decision depends crucially

on what firms expect the present discounted value of profits per hire to be; this

motivates the study of expectation formation at the firm level. In this paper, we relax the

assumption of rational expectations (RE) to study the role of adaptive **learning** (AL) on

job creation in the st**and**ard search **and** matching model. We show that the combination

of AL with simple forecasting models can match the volatility of US **labour** **market** **and**

survey data very well, outperforming the st**and**ard RE model.

In particular, under **learning** with autoregressive beliefs, such as AR(1), AR(2) **and**

VAR(1), we find that Total Factor Productivity (TFP) innovations can generate greater

incentive for vacancy creation than RE. A TFP innovation under **learning** increases the

firms’ present discounted value of profits per hire (i.e., firms become overoptimistic).

Thus, expectation formation has a central role to play in providing a solution to unemployment

volatility puzzle. Moreover, our simple **learning** set-up has the additional advantage

of being able to match the properties of forecast data on unemployment, which

is compatible with the expectation formation assumption in the paper.

34 **Staff** **Working** **Paper** **No**. **633** December 2016

References

Adam, K., **and** A. Marcet (2011): “Internal rationality, imperfect **market** knowledge

**and** asset prices,” Journal of Economic Theory, 146(3), 1224—1252.

Adam, K., A. Marcet, **and** J. Beutel (2016): “Stock Price Booms **and** Expected

Capital Gains,” mimeo., University of Mannheim.

Alves,S.A.L.(2012): “Trend Inflation **and** the Unemployment Volatility Puzzle,”

**Working** **Paper**s Series 277, Central Bank of Brazil, Research Department.

Andolfatto, D. (1996): “Business Cycles **and** Labor-Market Search,” American Economic

Review, 86(1), 112—32.

Barnichon, R. (2010): “Building a composite Help-Wanted Index,” Economics Letters,

109(3), 175—178.

Blanchard, O., **and** J. Gali (2010): “Labor Markets **and** Monetary Policy: A

New Keynesian Model with Unemployment,” American Economic Journal: Macroeconomics,

2(2), 1—30.

Branch, W. (2004): “Restricted Perceptions Equilibria **and** Learning in Macroeconomics,”

in Col**and**er (2004).

Branch, W. A., **and** G. W. Evans (2006a): “Intrinsic heterogeneity in expectation

formation,” Journal of Economic Theory, 127(1), 264—295.

(2006b): “A simple recursive forecasting model,” Economics Letters, 91(2),

158—166.

Bullard, J., **and** K. Mitra (2002): “Learning about monetary policy rules,” Journal

of Monetary Economics, 49(6), 1105—1129.

Campbell, J. Y. (1994): “Inspecting the mechanism: An analytical approach to the

stochastic growth model,” Journal of Monetary Economics, 33(3), 463—506.

Cho, I.-K., N. Williams, **and** T. J. Sargent (2002): “Escaping Nash Inflation,”

Review of Economic Studies, 69(1), 1—40.

Christiano, L., M. Eichenbaum, **and** M. Trab**and**t (2016): “Unemployment **and**

Business Cycles,” Econometrica, 84(4), 1523—1569.

Christoffel, K., **and** K. Kuester (2008): “Resuscitating the wage channel in models

with unemployment fluctuations,” Journal of Monetary Economics, 55(5), 865—887.

35 **Staff** **Working** **Paper** **No**. **633** December 2016

Cogley, T., **and** T. J. Sargent (2008): “Anticipated Utility And Rational Expectations

As Approximations Of Bayesian Decision Making,” International Economic

Review, 49(1), 185—221.

Col**and**er, D. (ed.) (2004): Post Walrasian Macroeconomics: Beyond the Dynamic

Stochastic General Equilibrium Model. Cambridge University Press, Cambridge.

Colciago, A., **and** L. Rossi (2011): “Endogenous Market Structures **and** Labor Market

Dynamics (New version),” Quaderni di Dipartimento 155, University of Pavia,

Department of Economics **and** Quantitative Methods.

Costain, J. S., **and** M. Reiter (2008): “Business cycles, unemployment insurance,

**and** the calibration of matching models,” Journal of Economic Dynamics **and** Control,

32(4), 1,120—55.

Davis, S., J. Haltiwanger, **and** S. Schuh (1996): Job creation **and** destruction. MIT

Press.

den Haan, W. J., G. Ramey, **and** J. Watson (2000): “Job destruction **and** the experiences

of displaced workers,” Carnegie-Rochester Conference Series on Public Policy,

52(1), 87—128.

Di Pace, F., K. Mitra, **and** S. Zhang (2014): “**Adaptive** Learning **and** Labour Market

Dynamics,” CDMA **Working** **Paper** Series 201408, Centre for Dynamic Macroeconomic

Analysis.

Ellison, M., **and** T. Yates (2007): “Escaping Volatile Inflation,” Journal of Money,

Credit **and** Banking, 39, 981—993.

Eusepi, S., **and** B. Preston (2011): “Expectations, Learning, **and** Business Cycle

Fluctuations,” American Economic Review, 101(6), 2844—72.

Evans, G., **and** S. Honkapohja (2001): Learning **and** Expectations in Macroeconomics

(Frontiers of Economic Research). Princeton.

Evans, G. W., **and** S. Honkapohja (2003): “Expectations **and** the Stability Problem

for Optimal Monetary Policies,” Review of Economic Studies, 70(4), 807—824.

(2006): “Monetary Policy, Expectations **and** Commitment,” Sc**and**inavian Journal

of Economics, 108, 15—38.

Evans, G. W., S. Honkapohja, **and** N. Williams (2010): “Generalized Stochastic

Gradient Learning,” International Economic Review, 51(1), 237—262.

36 **Staff** **Working** **Paper** **No**. **633** December 2016

Gertler, M., L. Sala, **and** A. Trigari (2008): “An estimated monetary DSGE model

with unemployment **and** staggered nominal wage bargaining,” Journal of Money, Credit

**and** Banking, 40(8), 1,713—64.

Gertler, M., **and** A. Trigari (2009): “Unemployment fluctuations with staggered

Nash wage bargaining,” Journal of Political Economy, 117(1), 38—86.

Gomes, P. (2011): “Fiscal policy **and** the **labour** **market**: the effects of public sector employment

**and** wages,” European Economy - Economic **Paper**s 439, Directorate General

Economic **and** Monetary Affairs (DG ECFIN), European Commission.

Guerrieri, V. (2008): “Heterogeneity, Job Creation **and** Unemployment Volatility,”

Sc**and**inavian Journal of Economics, 109(4), 667—693.

Haefke, C., M. Sonntag, **and** T. van Rens (2013): “Wage Rigidity **and** Job Creation,”

Journal of Monetary Economics, 60(8).

Hagedorn, M., **and** I. Manovskii (2008): “The cyclical behavior of equilibrium unemployment

**and** vacancies revisited,” American Economic Review, 98(4), 1,692—706.

Hall, R. E. (2005): “Employment fluctuations with equilibrium wage stickiness,” American

Economic Review, 95(1), 50—65.

Heemeijer, P., C. Hommes, J. Sonnemans, **and** J. Tuinstra (2009): “Price stability

**and** volatility in **market**s with positive **and** negative expectations feedback: An

experimental investigation,” Journal of Economic Dynamics **and** Control, 33(5), 1052—

1072.

Hertweck, M. S. (2013): “Strategic wage bargaining, labor **market** volatility, **and**

persistence,” The B.E. Journal of Macroeconomics, 13(1),27.

Hommes, C., J. Sonnemans, J. Tuinstra, **and** H.v**and**eVelden(2005): “Coordination

of Expectations in Asset Pricing Experiments,” Review of Financial Studies,

18(3), 955—980.

Hosios, A. J. (1990): “On the Efficiency of Matching **and** Related Models of Search **and**

Unemployment,” Review of Economic Studies, 57(2), 279—98.

Huang, K., Z. Liu, **and** T. Zha (2009): “Learning, **Adaptive** Expectations **and** Technology

Shocks,” Economic Journal, 119, 377—405.

Jacobs, D., E. Kalai, **and** M. Kamien (eds.) (1998): Frontiers of Research in Economic

Theory. Cambridge University Press, Cambridge.

37 **Staff** **Working** **Paper** **No**. **633** December 2016

Krause, M., **and** T. A. Lubik (2007): “Does intra-firm bargaining matter for business

cycle **dynamics**?,” Discussion paper.

Kreps, D. M. (1998): “Anticipated Utility **and** Dynamic Choice,” in Jacobs, Kalai, **and**

Kamien (1998), pp. 242—274.

Kuang, P., **and** K. Mitra (2016): “Long-Run Growth Uncertainty,” Journal of Monetary

Economics, 79, 67—80.

McGough, B. (2006): “Shocking Escapes,” Economic Journal, 116, 507—528.

Menzio, G. (2005): “High frequency wage rigidity,” Discussion paper, mimeo.

Menzio, G., **and** S. Shi (2011): “Efficient Search on the Job **and** the Business Cycle,”

JournalofPoliticalEconomy, 119(3), 468 — 510.

Milani, F. (2007): “Expectations, **learning** **and** macroeconomic persistence,” Journal of

Monetary Economics, 54(7), 2065—2082.

(2011): “Expectation Shocks **and** Learning as Drivers of the Business Cycle,”

Economic Journal, 121, 379—401.

Mitra,K.,G.W.Evans,**and** S. Honkapohja (2013): “Policy change **and** **learning**

in the RBC model,” Journal of Economic Dynamics **and** Control, 37(10), 1947—1971.

(2016): “Fiscal Policy Multipliers in an RBC Model with Learning,” Macroeconomic

Dynamics, forthcoming.

Mortensen, D., **and** E. Nagypal (2007): “More on Unemployment **and** Vacancy

Fluctuations,” Review of Economic Dynamics, 10(3), 327—347.

Mortensen, D. T., **and** C. A. Pissarides (1994): “Job Creation **and** Job Destruction

in the Theory of Unemployment,” Review of Economic Studies, 61(3), 397—415.

Ormeno, A., **and** K. Molnar (2015): “Using Survey Data of Inflation Expectations

in the Estimation of Learning **and** Rational Expectations Models,” Journal of Money,

Credit **and** Banking, forthcoming.

Orphanides, A., **and** J. C. Williams (2007): “Robust monetary policy with imperfect

knowledge,” Journal of Monetary Economics, 54(5), 1406—1435.

Petrongolo, B., **and** C. A. Pissarides (2001): “Looking into the black box: a survey

of the matching function,” Journal of Economic Literature, 39(2), 390—431.

Petrosky-Nadeau, N. (2013): “Credit, Vacancies **and** Unemployment Fluctuations,”

Review of Economic Dynamics, Forthcoming.

38 **Staff** **Working** **Paper** **No**. **633** December 2016

Petrosky-Nadeau, N., **and** E. Wasmer (2013): “The Cyclical Volatility of Labor

Markets under Frictional Financial Markets,” American Economic Journal: Macroeconomics,

5(1), 193—221.

Pissarides, C. A. (2009): “The unemployment volatility puzzle: is wage stickiness the

answer?,” Econometrica, 77, 1,339—369.

Preston, B. (2005): “Learning about Monetary Policy Rules when Long-Horizon Expectations

Matter,” International Journal of Central Banking, 1(2).

(2006): “**Adaptive** **learning**, forecast-based instrument rules **and** monetary policy,”

Journal of Monetary Economics, 53(3), 507—535.

(2008): “**Adaptive** **learning** **and** the use of forecasts in monetary policy,” Journal

of Economic Dynamics **and** Control, 32(11), 3661—3681.

Quadrini, V., **and** A. Trigari (2008): “Public Employment **and** the Business Cycle,”

Sc**and**inavian Journal of Economics, 109(4), 723—742.

Reiter, M. (2007): “Embodied technical change **and** the fluctuations of unemployment

**and** wages,” The Sc**and**inavian Journal of Economics, 109(4), 695—721.

Robin, J.-M. (2011): “On the Dynamics of Unemployment **and** Wage Distributions,”

Econometrica, 79(5), 1327—1355.

Sargent, T. J. (1999): The Conquest of American Inflation. Princeton University Press,

Princeton NJ.

Shimer, R. (2005): “The cyclical behavior of equilibrium unemployment **and** vacancies,”

American Economic Review, 95(1), 25—49.

(2010): Labor Markets **and** Business Cycles (Crei Lectures in Macroeconomics).

Princeton University Press.

Slobodyan, S., **and** R. Wouters (2012): “Learning in a Medium-Scale DSGE Model

with Expectations Based on Small Forecasting Models,” American Economic Journal:

Macroeconomics, 4(2), 65—101.

Trigari, A. (2006): “The Role of Search Frictions **and** Bargaining for Inflation Dynamics,”

**Working** **Paper**s 304, Innocenzo Gasparini Institute for Economic Research,

Bocconi University.

39 **Staff** **Working** **Paper** **No**. **633** December 2016

A

Appendix

A.1 Wage Bargaining

The surplus of workers in firm , W , is given by the marginal value of employment,

equation (46), expressed in terms of goods

W = H ( )

(46)

Thus,

W = − +[1− − ( )] −1

∗ W +1 (47)

The negotiated wage is set to maximise the joint surplus of a match between workers

**and** firm ,

arg max (W ) (V ) 1−

where ∈ (0 1) denotes the workers’ bargaining power or the share of the surplus the

worker is able to take. The first order condition of this problem is given by

(1 − ) W = V (48)

Substituting for equation (17) **and** (46) gives the negotiated wage

= (1− ) © − [1 − − ( )] −1

∗ W +1

ª

+

+ £ −1

+(1− ) −1

∗ V +1

¤

Because of continuous re-negotiation, we assume that the first order condition also holds

for subjective expectations

to get the following expression

Using (49) yields

(1 − ) ∗ W +1 = ∗ V +1 (49)

=(1− ) £ ¤

+ ( )

−1 ∗ W +1 + −1

(50)

=(1− ) + £ −1

+ ( ) −1

∗ V +1

¤

(51)

**and** substituting (16) gives

=(1− ) +

∙

¸

−1

+ ( ) (52)

( )

40 **Staff** **Working** **Paper** **No**. **633** December 2016

Rearranging gives equation (21) in the main text.

A.2 Behavioural Rules

In this appendix derive the linearised consumption **and** vacancy decision rules, equations

(24) **and** (25)

A.2.1

Job Creation Equation

We first linearise equations (15) **and** (19)

˜ +1 =(1− )˜ +¯˜ +¯˜

− ¯ ˜ 2 = − ¯¯−1 − ¯

˜

¯ 2 + (53)

1

¯ ¯( − 1)¯−2˜ +1 + 1¯ ¯−1˜ +1 − 1¯ ˜ +1 +

∞X

£

(1 − ) −1

∗ ¯( − 1)¯

−2˜ + + ¯ −1˜ ¤

+ − ˜ + −

=2

(¯¯ −1 − ¯)

∞X

X

(1 − ) −1 +1 ∗ ˜ +−1

=2

By replacing ˜ +1 into (53) **and** using =1¯, weobtain

∙ ¸

¯ 1¯˜ − ¯ + ¯ 1¯ ˜ 2 = ¯ 1 (1 − )˜ − (¯¯ 2 − ¯) 2 ∞X

1 − (1 − ) ˜ + (1 − ) −1 ∗¯ 1˜ + +

=2

"

∞X

(1 − ) −1

∗ ¯ 2˜ + − ˜ + − ¡¯¯ 2 − ¯ ¢ #

−1

X

˜ + (54)

=1

where ¯ 1 = ( − 1)¯¯ −2 **and** ¯ 2 = ¯ −1 A more compact representation gives (25) in

the main text

µ

¯ 1¯˜ − ¯ + ¯ 1¯ ˜ 2 = ¯ 1 (1 − )˜ − (¯¯ 2 − ¯) 2

1 − (1 − ) ˜ + S + S − S − S (55)

where

=1

=1

S =

S =

∞X

∞X

(1 − ) −1 ¯2 ∗ ˜ + S = (1 − ) −1 ¯1 ∗ ˜ +

=1

=2

−1

∞X

∞X

X

(1 − ) −1 ∗ ˜ + **and** S =(¯¯ 2 − ¯) (1 − ) −1 +1 ∗ ˜ + .

=1

=1

=1

41 **Staff** **Working** **Paper** **No**. **633** December 2016

A.2.2

Consumption Function

By linearising (8) **and** substituting into the linearised version of equation (10), we get the

following consumption rule

1

1 − ˜ (¯¯ +¯)2

= ˜ +¯ ˜ +¯˜ + ¯˜ +1 − ˜

1 −

∞X

∞X

+ ∗ ¯ ˜ + + ∗ ¯˜ + − (¯¯ +¯) 2

where

≡

=1

=2

˜ +¯ ˜ +¯˜ + ¯˜ +1 −

∞

X

=1

"

#

−1

X

∗ ˜ + +

=1

(¯¯ +¯)2

˜ + S + S + S − S

1 −

∞X

∗ ˜ +

=1

S =

S =

∞X

∞X

¯ ∗ ˜ + S = ¯ ∗ ˜ +

=1

=2

"

∞X

∞X

∗ ˜ + **and** S =(¯¯ +¯) +1

∗

=1

=1

#

X

˜ +

=1

A.3 Learning

For expositional purpuses, we derive the infinite sums under AR(2). The derivation of the

infinite sums under AR(1) **and**VAR(1) belief specifications are available upon request.

As noted in the main text, the AR(2) belief specification is given by

= 0 + 1 −1 + 2 −2 for = { +1 }

³

³

Let Φ = 0 1 2´0

**and** Ψ = 1 −1 −2´0.

Agents use a constant gain

Recursive Least Squares (RLS) algorithm to update their beliefs

Φ = Φ −1 + Ψ −1

−1

³ −1 − Φ 0 −1Ψ −1´0

(56)

= −1 +

³Ψ −1 Ψ 0 −1 − −1´

where ∈ (0 1) denotes the constant gain **learning** parameter **and** is the precision (3

x 3) matrix associated with each equation. In addition, we assume that agents estimate

the persistence parameter of the exogenous productivity process (12) using constant

gain RLS; this captures their uncertainty about the persistence of the productivity process

perhaps changing over time.

42 **Staff** **Working** **Paper** **No**. **633** December 2016

A.3.1

Computation of Infinite Sums

Since the AR(2) forecasting models are estimated in levels, agents’ perceptions of the

steady state changes with the arrival of new information. A bar over the variable together

with the superscript denotes the changes in the perception of the steady state at each

point in time. Under AR(2), the perceived steady states are

¯ 0

=

for = { }

1 − 1 − 2

Let’s denote the deviations of the variable from the perceived steady state with a

tilde **and** a superscript .

˜ = − ¯ for = { +1 }

Knowing the evolution of the perceived steady states, the belief system can then be

written as

˜ = 1˜ −1 + 2˜ −2 for = © +1

ª

**and** = { }

Finally, we can express each equation in VAR(1) form, ˜ = Λ ˜

Ã !

−1 , where

³ ´0

˜

= ˜ ˜ 1 2

−1

**and** Λ =

. Thus, it follows that future forecasts

1 0

³ ´

can simply be computed as ˜ + = 1 0 Λ +1 ˜ −1 for = { } **and** ˜ =

ª

©˜ +1 ˜ ˜ ˜

A.3.2

Computational Details of Sums

The infinite sums under ³ AR(2) ´ beliefs are derived using the decomposition ˜ =( − ¯ )+

(¯ − ¯) **and** ˜ + = 1 0 Λ +1

˜ −1 for = { } **and** ˜ = ©˜ ª

+1 ˜ ˜ ˜ .

S , S **and** S in equation (25) are computed as

S =

∞X

(1 − ) −1 ¯2 ∗ ˜ + =

=1

∞X

(1 − ) −1 ¯2ˆ ˜ =

=1

ˆ¯ 2

1 − (1 − )ˆ ˜

43 **Staff** **Working** **Paper** **No**. **633** December 2016

S =

∞X

(1 − ) −1 ∗ ˜ + =

=1

∞X

(1 − ) −1 ˜ + +

∞X

= (1 − ) −1 ˜ + + (¯ − ¯)

1 − (1 − )

=1

∞X

´

= (1 − ) −1

³1 0 Λ +1 ˜ −1 + (¯ − ¯)

1 − (1 − )

=1

³ ´

= 1 0 Λ 2 [ − (1 − ) Λ ] ˜ −1 +

(¯ − ¯)

1 − (1 − )

=1

∞X

(1 − ) −1 (¯ − ¯)

=1

S =

=

=

∞X

(1 − ) −1 ¯1 ∗ ˜ + =

=2

∞X

(1 − ) −1 ˜ + +

=2

∞X

(1 − ) −1 (¯ − ¯)

=2

∞X

(1 − ) −1 ˜ + + ¯ 1 2 (1 − )(¯ − ¯)

1 − (1 − )

∞X

´

(1 − ) −1

³1 0 Λ ˜ −1 + ¯ 1 2 (1 − )(¯ − ¯)

1 − (1 − )

=2

=2

= 2 (1 − )

³ ´

1 0 Λ 2 [ − (1 − ) Λ ] ˜ −1 + ¯ 1 2 (1 − )(¯ − ¯)

1 − (1 − )

The sum S contains a double summation operator **and** it requires using P ∞

=1 (1 −

) = (1−) The computation of S [1−(1−)] 2 is thus

44 **Staff** **Working** **Paper** **No**. **633** December 2016

∞

"

#

−1

X

X

S = (¯¯ 2 − ¯) 2 (1 − ) ∗ ˜ +

=1

= (¯¯ 2 − ¯) 2 ( ∞

X

=1

=1

h

i

(1 − ) ( − ¯)

= (¯¯ 2 − ¯) 2 (

(1 − )( − ¯)

[1 − (1 − )] 2 +

= (¯¯ 2 − ¯) 2 (

(1 − )( − ¯)

[1 − (1 − )] 2 +

+

"

#)

∞X

X−1

(1 − ) ∗ ˜ +

=1

=1

"

#)

−1

∞X

X

(1 − ) ∗ ˜ +

=1

=1

=1

"

#)

−1

∞X

X ³ ´

(1 − ) 1 0 Λ +1

˜ −1

= (¯¯ 2 − ¯) 3 (1 − )( − ¯)

+

[1 − (1 − )] 2

³ ´

∞X h

(¯¯ 2 − ¯) 3 (1 − ) 1 0 ( − Λ ) −1 Λ 2 (1 − ) −1 ¡ ¢ −1 − Λ −1

˜ −1i

= (¯¯ 2 − ¯) 3 (1 − )( − ¯)

+

[1 − (1 − )]

³

2 ´

(¯¯ 2 − ¯) 3 (1 − ) 1 0 ( − Λ ) −1

h ¡

(1 − (1 − ) ) −1 − Λ ( − (1 − ) Λ ) −1¢ ˜ −1i

Λ 2

A similar approach is used to compute the infinite sums in the household’s consumption

behavioural rule. S , S **and** S in equation (24) are calculated as follows

S = ¯

"

∞X

X ∞

∗ ˜ + =¯ ˜ + +

=1

=1

=1

=1

#

∞X

(¯ − ¯)

=1

" ∞

# "

X

= ¯ ˜ + + (¯ − ¯) X ∞

=¯

1 −

=1

=1

∙ ³ ´

= ¯ 1 0

Λ 2 ( − Λ ) −1 ˜ −1 + (¯ − ¯)

1 −

´

³1 0 Λ +1 ˜ −1 + (¯ − ¯)

1 −

¸

#

( ∞

) (

X

X ∞

S =

∗ ˜ + =

∗ ˜ + +

=1

=1

∞X

= ˜ + + (¯ − ¯)

=

1 −

=1

³ ´

= 1 0

³

1 0

´ ∞

X

=1

)

∞X

(¯ − ¯)

=1

−1 Λ +1

Λ 2 ( − Λ ) −1 ˜ −1 + (¯ − ¯)

1 −

˜ −1 + (¯ − ¯)

1 −

=

45 **Staff** **Working** **Paper** **No**. **633** December 2016

S = ¯

= ¯

= ¯

= ¯

"

∞X

X ∞

∗ ˜ + =¯ ˜ + +

=2

" ∞

X

=2

" ∞

X

=2

=2

˜ + + 2 (¯ − ¯)

1 −

#

#

∞X

(¯ − ¯)

=2

´

³1 0 Λ ˜ −1 + 2 (¯ − ¯)

1 −

∙ 2 ³

1 0

´

#

=

Λ 2 ( − Λ ) −1 ˜ −1 + 2 (¯ − ¯)

1 −

¸

The sum S contains a double summation operator **and** it involves using P ∞

=1 =

(1−) 2 .Thus,

∞

"

X

S = (¯¯ +¯) 2

=1

#

X

∗ ˜ +

−1

=1

(

X ∞

£

= (¯¯ +¯) 2 ( − ¯) ¤ +

=1

= (¯¯ +¯) 2 (

( − ¯)

1 − 2 +

= (¯¯ +¯) 2 (

( − ¯)

1 − 2 +

"

∞X

=1

=1

"

∞X

=1

X

−1

=1

#)

X

∗ ˜ +

−1

=1

³

1 0

½ (

= (¯¯ +¯) 3 − ¯)

³ ´

1 − 2 + 1 0 ( − Λ ) −1 Λ 2

´

Λ +1

˜ −1

#)

∞X ³ ´ ¡Λ ¢ i h )

−1 2

1 0 − Λ +2

( − Λ ) −1 ˜ −1

£

(1 − ) −1 − Λ ( − Λ ) −1¤ ˜ −1

¾

46 **Staff** **Working** **Paper** **No**. **633** December 2016

B

Additional Tables

Recall that the difference between the RE model (36) **and** the AL model with correctly

specified beliefs (37) is that, whilst under RE agents know the parameters of the RE

solution, under **learning** agents do not know the value of the parameters **and** update

them as new information becomes available. Tables 8 **and** 9 below show that this model

suffers from the same problem as the RE model i.e. it is not able to provide a solution

to the unemployment volatility puzzle **and** unable to match the forecast error properties

in the data.

Model Statistics ˆ ˆ ˆ ˆ

ˆ ˆ

ˆ1 ˆ 1 003 050 043 091 100

(b 1 b 1−1 ) 096 097 092 097 096 096

ˆ 1 099 099 −099 100 100

RE ˆ − 1 096 −100 099 099

(b 1 b 2 ) ˆ − − 1 −096 099 099

ˆ − − − 1 −099 −099

ˆ − − − − 1 100

ˆ − − − − − 1

ˆ1 ˆ 1 003 050 043 092 100

(b 1 b 1−1 ) 096 097 092 097 096 096

AL with ˆ 1 100 099 −099 100 100

correctly ˆ − 1 098 −100 100 100

specified (b 1 b 2 ) ˆ − − 1 −096 099 099

beliefs ˆ − − − 1 −099 −099

ˆ − − − − 1 100

ˆ − − − − − 1

Table 8: Summary Statistics: RE **and** Correctly Specified Beliefs

**No**tes. Relative st**and**ard deviations, autocorrelation **and** correlation coefficients in this Table correspond to the quaterly

simulated series expressed in percentage deviations from steady state value. The term ( 1 2 ) st**and**s for the correlation

coefficient between variables 1 **and** 2

1

2

3

4

ˆ1 ˆ 005 012 018 021

RE (ˆ 1 ˆ 1−1 ) 013 024 061 073

(ˆ 1 ∆ˆ ) −010 −004 −003 000

(ˆ 1 ∆ˆ ) 027 000 004 000

AL with ˆ1 ˆ 004 012 018 021

correctly (ˆ 1 ˆ 1−1 ) −006 024 061 073

specified (ˆ 1 ∆ˆ ) −001 −004 −004 −002

beliefs (ˆ 1 ∆ˆ ) −002 000 004 001

Table 9: Forecast Properties of RE **and** Correctly Specified Beliefs

**No**tes. The term ( 1 2 ) st**and**s for the correlation coefficient between variables 1 **and** 2 . Data from Survey of

Professional Forecasters. Forecast errors

for = {1 2 3 4} as defined in Table 5

47 **Staff** **Working** **Paper** **No**. **633** December 2016

The table below reports on the forecast error properties for the robustness exercise in

Section 6..

1

2

3

4

ˆ1 ˆ 456 621 802 989

Data (ˆ 1 ˆ 1−1 ) 070 079 085 089

(ˆ 1 ∆ˆ ) −069 −055 −047 −036

(ˆ 1 ∆ˆ ) 044 042 037 032

Alternative parameterisations

ˆ1 ˆ 308 464 565 623

AR(2) ( =0001) (ˆ 1 ˆ 1−1 ) 048 082 089 092

(ˆ 1 ∆ˆ ) −008 −016 −015 −017

(ˆ 1 ∆ˆ ) 015 021 020 022

ˆ1 ˆ 333 465 540 596

AR(2) ( =0003) (ˆ 1 ˆ 1−1 ) 019 058 075 081

(ˆ 1 ∆ˆ ) −015 −011 −012 −012

(ˆ 1 ∆ˆ ) 020 018 016 016

ˆ1 ˆ 349 465 537 586

AR(2) ( =07) (ˆ 1 ˆ 1−1 ) 018 057 072 079

(ˆ 1 ∆ˆ ) −011 −010 −011 −008

(ˆ 1 ∆ˆ ) 015 013 013 011

ˆ1 ˆ 453 578 657 706

AR(2) ( =06 **and** =04) (ˆ 1 ˆ 1−1 ) 019 061 074 079

(ˆ 1 ∆ˆ ) 000 −007 −010 −011

(ˆ 1 ∆ˆ ) 004 010 012 014

Table 10: Forecast Properties under Alternative Parameterisations

**No**tes. Data from Survey of Professional Forecasters. Relative st**and**ard deviations in this Table correspond to the

quarterly simulated series expressed in percentage deviations from steady state values under different parameterisations.

The term ( 1 2 ) st**and**s for the correlation coefficient between variables 1 **and** 2 . Forecast errors

= {1 2 3 4} as defined in Table 5.

for

48 **Staff** **Working** **Paper** **No**. **633** December 2016