real exchange rate behavior: new evidence from …...real exchange rate behavior: new evidence from...

Real Exchange Rate Behavior:

New Evidence from Matched Retail Prices

Alberto Cavallo

MIT and NBER

Preliminary

Brent Neiman

U. Of Chicago and NBER

Roberto Rigobon

MIT and NBER

Classic question: how are tradable goods� prices and exchange rates related?

A vast literature attempts to characterize the level and behavior of the the Real Exchange Rate

Common finding � little co-movement between relative prices in local currencies (RP) and nominal exchange rates (NER), particularly at the retail level

The RER co-moves with the NER

Passthrough is low (Goldberg and Knetter (97), Burstein and Gopinath (2011))

RER shocks are persistent � PPP Puzzle (Rogoff (96))

RER for tradeables is just as volatile as for non-tradeables (Engel (99))

Motivation

Standard explanation would be Balassa-Samuelson-type effects. But, hard to find evidence that traded good RER behaves so differently (Engel 1999).

Other possibilities include:

Distribution costs (Burstein et al 2003)

Selection into trade and variable markups (Atkeson and Burstein 2008)

Sectoral aggregation more problematic in CPI (Imbs et al 2005, Carvalho and Nechio 2011)

Biases from temporal aggregation more problematic in CPI (Taylor 2001)

Biases from disregard of entering and exiting goods (Nakamura and Steinsson 2012)

Motivation

Most papers use CPIs/PPIs/IPIs

Not designed for international comparisons

Different index methods and goods across countries

No levels, no entry/exit

Existing micro-data limited to few countries (e.g. US-Canada), goods (eg.

Big Mac index), and/or degree of product matching (eg. EIU).

World Bank’s International Comparisons Program comes close but among

other issues is extremely low frequency.

Summarized Nicely by Taylor (ECMA 2001):

Huge Empirical Challenge

“To meet the desired standard we would be hoping that hundreds of price inspectors

would leave a hundred or more capital cities on the final day of each month, scour

every market in all representative locations, for all goods, and come back at the end

of a very long day, with a synchronized set of observations from Seoul to Santiago,

from Vancouver to Vanuatu. We cannot pretend that this happens.”

We use web-scraping methods to substitute for Taylor’s hundreds of price

inspectors” and achieve:

Better matching of products and methods (relative to CPIs)

Higher frequency (relative to ICP)

Match more than 50,000 individual goods (“varieties”) to approx. 350 narrow

product definitions (“products”) chosen to fully cover the CPI categories of

food, fuel, and electronics at the retail level in 9 countries from (some subset

of) 2010-2016.

Evaluate traded good real exchange rate behavior (bilaterally with U.S.) in our

data and compare with CPI-based equivalents

What We Do

Regress relative prices on exchange rates and find roughly 0.75 with our

matched retail price data compared to 0.34 for equivalent measures in the

tradable CPI data (same countries, sectors, and time period)

Source of differences mostly be fact that we exactly match prices together with

a margin from price levels at entry/exit of varieties

What We Find

International Finance

PPP Puzzle Debate: Rogoff (96) Imbs et al. (QJE 2005), Chen and Engel (PER 2005), Carvalho

and Nechio (AER 2011)

Decomposing RER Adjustment, Macro and Micro: Bergin et al. (ReStat 2013), Crucini and

Shintani (JME 2008)

Extensive Margin and Price Indices: Nakamura and Steinsson (2012), Gagnon, Mandel, and

Vigfusson (2014)

RER Behavior Using Online Prices: CNR (2014, 2015), Gorodnichenko and Talavera (2014)

Passthrough Goldberg and Knetter (97), Engel (99), Gopinath, Itskhoki and Rigobon (2007),

Burstein and Gopinath (2011)

International Comparisons

Penn World Tables: Heston and Summers (88,96), Nuxoll (94), Feenstra, Inklaar and Timmer

(2013)

Measuring PPPs: Diewert (99), Hill(99), Deaton and Heston (10)

Poverty and PPPs: Deaton(06), Deaton (10)

ICP Rounds and “Surprises”: Diewert(10, 13), World Bank (14), Inklaar and Rao (16), Deaton

and Aten (16)

Overview of Related Literature

Private company linked to the Billion Prices Project

Use web-scraping methods to collect daily data for all goods sold by thousands of

large retailers in 50 countries

Unlike CNR (2014, 2015), difficult to find goods in broad categories that are truly

identical. Barcodes, package sizes, flavors, brands, etc., vary by country.

We follow the ICP methodology:

Create a list of narrow product definitions � a “product”

Each individual good/UPC we find at the store is a “variety”

Match varieties to each product

Look within largest retailers (typically 3-5, but varies) for matches, code

algorithm to generate unit values where needed

Data from PriceStats

Classification System

“Product” level

UN’s Classification of Individual

Consumption According to Purpose

Standard

COICOP

Levels 1-4

New

Levels

CPI weights only

available up to Level 3

(or 1 in some

countries)

• Products chosen to have good category coverage, also found in many countries

Product Definitions

Convert to Unit Prices, Calculate product-level RER

Ground Coffee - Regular

Adjusting for size

Step 1: Unit Prices for all varieties

in each country

Prices need to be adjusted by package size (which can be extremely large in the US)

Product Varieties (Size)

Convert to Unit Prices, Calculate product-level RER

The price for product i, country y, time t is:

Ground Coffee – Regular – 1 gram

where is the set of varieties captured by

our scraping.

Average Price US (dollars) Average Price UK (pounds)

Step 2: Average unit prices for all

products in each country

Methodology

Bilateral Product RER

Ground Coffee – Regular – 1 gram

The nominal exchange rate is expressed as units of z per unit of y

(increase is appreciation of y)

Step 3: Form relative prices,

combine with NER,

form RER

Matched Dataset

2011 Comparison with World Bank’s ICP at 3-digit level

2011 Comparison with World Bank’s ICP at 1-digit level

Decomposing the Relationship Between our RERs and CPI-

Based Ones

Step 1: Generate RER using All Item CPIs

Step 2: Only use 1-Digit CPIs corresponding to our PPP dataset

Step 3: Only use the 3-Digit CPIs within those 1-digits the correspond to our

PPP dataset

Step 4: Create the 3-digit CPI-based RERs using Fisher weights, then aggregate

Step 5: Do this for PPP Data, but only using continuing goods (i.e. like a Matched

Model)

Step 6: Do this for all goods, including entries/exits

Our RERs and CPI-Based RERs

Literature employs many econometric methods:

Correlations and Granger Causality

AR1s and Stationarity Tests

Vector Error Correction Models

Have tried some with potential for useful results, but difficult

given short time-series.

We run regressions of relative price on exchange rate (ala

Gorodnichenko and Talavera, AER 2016):

Always run bilaterally with US

β = - 1 implies fixed RER

Note that these are relative passthrough coefficients.

Beyond Visual Evidence

Relative Passthrough is 75% on average

Our RERs and CPI-Based RERs: Step by Step

CPI-based

Goods CPI or manufacturing PPI often used as rough proxies for

tradables (eg. Engel (99))

We do a bit better:

1-digit CPI: uses official CPIs for level 1 sectors in our data:

Food and Beverages

Transportation

Recreation and Culture

Sectoral differences


Same Sector

(food, fuel, elect)

Goods CPI or manufacturing PPI often used as rough proxies for

tradables (eg. Engel (99))

We do a bit better:

1-digit CPI: uses official CPIs for level 1 sectors in our data:

Food and Beverages

Transportation

Recreation and Culture

3-digit CPI: uses level 3 official CPIs (where available), to build an

aggregate CPI that excludes sub-sectors that are not in our data �

eg non-tradable services, eg. public transport, packaged holidays

Sectoral differences


Same Sector

(food, fuel, elect)

Same subsectors

Following the international comparisons (ICP) literature, we measure

RERs using Fisher indices that start with 3-digit relative prices and

aggregate up using expenditure weights from both countries.

When computing RERs with CPIs, the standard procedure is to first

get the aggregate CPI in each country separately, and then compute

the ratio.

This can cause a formula or “extrapolation bias”(see Deaton (2012)

and Inklaar and Rao (2016)):

Formula or “Extrapolation” Bias


Formula bias

Matched Data

Row (5) is an index that uses matched data (but does not yet

incorporate the effect of entry and exit price levels)

Passthrough rises 26 percentage points

Why?

Below discuss some possibilities and rule out others


PPP Data (matched

products)

Differences in Online and Offline Prices? Not likely.

Source: Cavallo (2017) “Are Online and Offline Prices Similar? Evidence from Large Multi-Channel Retailers”, American Economic Review Vol 107(1)

Difference in stickiness vs CPI data? Not really…

Our PPP data is actually stickier than comparable in US CPI data (food drives

this result)

Cavallo (2016) “Scraped Data and Sticky Prices” Review of Economics and

Statsitics � CPI imputations and scanner data averaging can increase the

frequency of price changes.

Matched products more flexible than other online data

Our hypothesis: Matching procedure “picks” particularly tradable goods

Imagine that true passthrough is one, but imperfect matching means we

use a proxy price which differs by some error term

If error term is correlated with the exchange rate, changes estimates:

Highly plausible that proxy uses different imported inputs, different

currency of invoicing, affected by global shocks

What could cause lower passthrough?

If CPIs include products with less imported inputs

If CPIs favor products with local currency of invoicing

Product Matching and Error Bias

Product Entry/Exit Bias

Ignoring a constant term, we can write t+1 prices as a function of

new and continuing prices

Alternatively, we can focus on the decomposition at time t (instead

of t+1) between exiting and continuing goods

Share of intros Price of intros Share of continuing Price of continuing


Combining these equations we obtain

where the first term is the matched-model index

Combining this equation across countries and products gives:


Product Intro

and exit Particularly

important for

electronics


Surprising that entry/exit appears to matter so much.

On one hand, plausible measurement and scraping issues might

accentuate margin (e.g. goods disappearing and re-entering

seasonally, with price changes).

On other hand, note that entry/exit huge for electronics, moderate

for food, zero for fuel. This seems economically sensible.

We build a high-frequency dataset of closely-matched products in 9

countries � construct RER, RP time series

We find strong co-movement of RP and E (not seen with CPIs)

We quantify sources of measurement bias in passthrough estimates:

Sectoral differences & formula [4 % points]

Product matching bias (matched goods, relative prices) [26 % points]

Product entry/exit bias [11 % points]

Summary

real exchange rate behavior: new evidence from …...real exchange rate behavior: new evidence from...

Documents