使用Google趨勢跟蹤實時經(jīng)濟(jì)活動

1、Table of contents HYPERLINK l _bookmark0 Tracking activity in real time with Google Trends HYPERLINK l _bookmark0 6 HYPERLINK l _bookmark1 1. Introduction and Summary HYPERLINK l _bookmark1 6 HYPERLINK l _bookmark3 2. The COVID-19 crisis called for the use of high-frequency indicators HYPERLINK l _b

2、ookmark3 7 HYPERLINK l _bookmark5 3. Exploiting the full potential of Google Trends HYPERLINK l _bookmark5 9 HYPERLINK l _bookmark10 4. A neural panel model of GDP growth HYPERLINK l _bookmark10 11 HYPERLINK l _bookmark11 4.1. From quarterly GDP growth to a weekly tracker: a bridge model of GDP grow

3、th HYPERLINK l _bookmark11 12 HYPERLINK l _bookmark12 4.2. A non-linear algorithm HYPERLINK l _bookmark12 12 HYPERLINK l _bookmark17 4.3. To pool or not to pool: a panel nowcasting model for 46 countries HYPERLINK l _bookmark17 14 HYPERLINK l _bookmark21 5. How well can the OECD Weekly Tracker nowca

4、st the economy? HYPERLINK l _bookmark21 15 HYPERLINK l _bookmark27 6. Model insights: a dive into the black box HYPERLINK l _bookmark27 20 HYPERLINK l _bookmark28 6.1. Shapley values: explaining machine learning with game theory HYPERLINK l _bookmark28 20 HYPERLINK l _bookmark29 6.2. A dive into the

5、 model inner workings HYPERLINK l _bookmark29 21 HYPERLINK l _bookmark35 6.3. From Shapley values to sectoral insights HYPERLINK l _bookmark35 23 HYPERLINK l _bookmark37 7. The OECD Weekly Tracker HYPERLINK l _bookmark37 23 HYPERLINK l _bookmark39 7.1. The COVID-19 crisis: a week-by-week analysis HY

6、PERLINK l _bookmark39 24 HYPERLINK l _bookmark42 7.2. Latest insights from the Weekly Tracker: a stalling recovery below 2019 levels HYPERLINK l _bookmark42 30 HYPERLINK l _bookmark44 7.3. Consumption volume remains subdued while its composition has shifted HYPERLINK l _bookmark44 31 HYPERLINK l _bo

7、okmark48 References HYPERLINK l _bookmark48 35 HYPERLINK l _bookmark49 Data pre-processing and data issues HYPERLINK l _bookmark49 40 HYPERLINK l _bookmark52 Additional details HYPERLINK l _bookmark52 44 HYPERLINK l _bookmark53 Additional results HYPERLINK l _bookmark53 47Tables HYPERLINK l _bookmar

8、k4 Table 1. Standard indicators were outpaced by the crisis HYPERLINK l _bookmark4 8 HYPERLINK l _bookmark25 Table 2. Forecast performance HYPERLINK l _bookmark25 18Figures HYPERLINK l _bookmark9 Figure 1. Queries in Google Trends: beyond keyword searches HYPERLINK l _bookmark9 11 HYPERLINK l _bookm

9、ark22 Figure 2. Nowcasting GDP growth with Google trends (M-1 forecast) HYPERLINK l _bookmark22 15 HYPERLINK l _bookmark24 Figure 3. Trackers predictions for Q2 2020 HYPERLINK l _bookmark24 17 HYPERLINK l _bookmark26 Figure 4. The OECD Weekly Tracker and Google Mobility HYPERLINK l _bookmark26 19 HY

10、PERLINK l _bookmark33 Figure 5. Most important variables and their contributions to predictions HYPERLINK l _bookmark33 22 HYPERLINK l _bookmark34 Figure 6. Partial dependence plots HYPERLINK l _bookmark34 22 HYPERLINK l _bookmark36 Figure 7. A focus on March 2020 23 HYPERLINK l _bookmark40 Figure 8

11、. Weekly Tracker: advanced economies 25 HYPERLINK l _bookmark41 Figure 9. Weekly Tracker: selected emerging economies 28 HYPERLINK l _bookmark43 Figure 10. Most recent predictions of the OECD Weekly Tracker 31 HYPERLINK l _bookmark45 Figure 11. Drivers of the recovery: aggregated Shapley Values 32 H

12、YPERLINK l _bookmark46 Figure 12. Consumption has decreased overall and shifted towards new patterns 33 HYPERLINK l _bookmark47 Figure 13. Google search intensities per spending categories 34Boxes HYPERLINK l _bookmark16 Box 1. Training the neural network 13 HYPERLINK l _bookmark18 Box 2. Sparse or

13、dense? 14Tracking activity in real time with Google TrendsNicolas Woloszko HYPERLINK l _bookmark2 11. Introduction and SummaryA pre-requisite for good macroeconomic policymaking is timely information on the current state of the economy, particularly when economic activity is changing rapidly. Given

14、that GDP is usually only available on a quarterly basis (with first estimates typically published only 4 weeks or more after the end of the quarter), policymakers and forecasters have long made use of more timely higher frequency data, such as survey-based indicators like Purchasing Managers Indices

15、 (PMIs). However, both the current crisis and the earlier ones have shown that the underlying relationship with survey-based indicators can become unreliable when changes in economic activity are abrupt and massive (Vermeulen, 20121).This problem has prompted a search for alternative high-frequency

16、indicators of economic activity. This paper discusses one such indicator based on Google Trends, which are used to construct a Weekly Tracker that provides real-time estimates of GDP growth in 46 G20, OECD and partner countries.The OECD Weekly Tracker of GDP growth attempts to fill the gap in real-t

17、ime high-frequency indicators of activity with a large country coverage. To the authors knowledge, the Tracker is the first weekly GDP proxy that can be compared across a large array of OECD and G20 countries. The Tracker provides estimates of year-on-year growth rate in weekly GDP with a 5-day dela

18、y. It applies a single machine learning algorithm on a panel of Google Trends data for 46 countries. The algorithm flexibly captures cross-country heterogeneity and provides comparable estimates across countries. It exploits the full potential of Google Trends data by aggregating together informatio

19、n about search behaviour related to consumption, labour markets, housing, industrial activity and uncertainty. The Tracker provides high- frequency and real-time information about the COVID-19 crisis and subsequent rebound, as well as the impact of confinement measures.The Tracker uses a two-step mo

20、del to nowcast weekly GDP growth based on Google Trends. First, a quarterly model of GDP growth is estimated based on Google Trends search intensities at a quarterly frequency. Second, the relationship between Google Trends and activity, using the same elasticities estimated from the quarterly model

21、, is applied to the weekly Google Trends series to yield a weekly tracker. The relationship between Google Trends variables and GDP growth is fitted using a machine learning algorithm (“neural network”). The algorithm captures non-linearities that are likely to beThe author is a member of the OECD E

22、conomics Department Macroeconomic Analysis Division and NAEC Innovation LAB, and thanks David Turner, Sebastian Barnes, Annabelle Mourougane, Boris Cournde, Dorothe Rouzet, Laurence Boone, Luiz de Mello, Alain de Serres, Nigel Pain, Sylvain Benoit, Frdric Gonzales, Daniela Glocker, Tim Bulman, Herms

23、 Morgavi, Sebastien Turban, Ana Chico Sanchez, and Vronica Humi, as well as participants of the Bank of England and Banca dItalia 2020 conferences on economics and machine learning. The author also thanks Hal Varian for granting the NAEC Innovation Lab access to the Google Trends API.key in extreme

24、situations, but which are difficult to estimate with more conventional econometric approaches.Using modern machine learning interpretability tools, this paper exploits the neural network to derive insights about non-linear patterns captured by the model that are consistent with economic intuition. F

25、or instance, searches for unemployment benefits are stronger predictors of activity around times when lay-offs increase and thus become dominant with regards to hiring in explaining changes in employment. Model interpretability tools also highlight the most important variables and the macroeconomic

26、predictive power of a number of topics including “bankruptcies”, “economic crisis”, “investment”, “l(fā)uggage” and “mortgage”.The model of GDP growth based on Google Trends proves to perform well in out-of-sample nowcast simulations. On average across OECD and G20 countries, the quarterly model based o

27、n Google Trends has a Root Mean Squared Error (RMSE) lower by 17% than an auto-regressive model that just uses lags of year-on-year GDP growth. It captures a sizeable share of business cycle variations, including around the Global Financial Crisis (when the available data for training the algorithm

28、was much smaller) and the euro area sovereign debt crisis. The timing of the downturn and subsequent rebound is well captured by the model, although the full magnitude of the negative shock in the second quarter of 2020 is typically under-estimated, given its unprecedented scale. The tracker thus pr

29、ovides a useful tool for real-time narrative analysis on a weekly basis, although it does not on average outperform models based on more standard variables, once these are eventually released. It also provides evidence of lasting shifts in consumption patterns away from services implying social inte

30、ractions.The paper is organised as follows. The second section describes the Google Trends data, data issues and data pre-processing. The third section introduces the non-linear modelling approach. The fourth section displays results of pseudo-real time simulations. The fifth section provides insigh

31、ts into the inner workings of the model using interpretability tools. And the sixth one shows the Weekly Tracker and provides insights on the 2020 recession.2. The COVID-19 crisis called for the use of high-frequency indicatorsThe 2020 crisis is unique in its magnitude and speed, and highlighted the

32、 caveats of standard indicators. Leading indicators most commonly used by policymakers fall in two categories: “hard” and “soft” (Table 1). Hard indicators are collected by national administrations or statistical agencies and suffer from publication delays ranging from one to three months, which is

33、a major constraint for policymakers facing rapid fluctuations in activity. Soft indicators are timelier, but can become less informative about GDP during recessions. PMIs and confidence surveys are often based on averages of qualitative answers based on the net balance of respondents optimism or pes

34、simism, which limits their ability to quantify the magnitude of an ongoing crisis.Table 1. Standard indicators were outpaced by the crisisIndicatorTypeFrequencyReleaseRelationship to GDPGDPHardQuarterly (monthly for GBR, CAN and SWE)Usually 1-2 months after the end of the quarterIndustrial productio

35、nHardMonthlyAround 30-55 days after the end of the monthLinearRetail salesHardMonthlyAround 8-10 weeks after the end of the monthLinearPMIsSoftMonthlyAround start of the next monthLinear in normal times, non-linear around crisesConsumer confidenceSoftMonthlyAround start of the next monthLinear in no

36、rmal times, non-linear around crisesGoogle MobilityHigh- frequencyDailyWith a 7-day delayDifficult to calibrate as historical data start mid-February 2020.Google TrendsHigh- frequencyDaily, Weekly or MonthlyWith a 5-day delayModel-based relationshipSource: OECD.As a specific example, the information

37、 provided by standard indicators to French policymakers when they implemented the lockdown in mid-March illustrates the limitations of these traditional gauges at a time of crisis. After the lockdown was implemented on 17 March, the first releases were the flash PMIs on 24 March, which sent mixed si

38、gnals reflecting the uneven nature of the shock as the manufacturing PMI fell moderately (to 42.9) while the services PMI fell to an all-time low (29.0). On 27 March, consumer confidence readings for February edged down marginally (to 103 from 104), well above market expectations (of 92), consistent

39、 with the unexpectedly high business confidence released one day before. Flash GDP releases for the first quarter came out on 30 April at -5.8% quarter-on-quarter, which did not provide specific information about activity in March as the GDP figure is a quarterly average. The first traditional hard

40、indicators to provide information about activity in March were household consumption (- 17.9% month-on-month) and industrial production (-16.2% month-on-month), but these were only published on 30 April and 7 May, respectively, over six weeks after the start of the lockdown.The need to quickly under

41、stand the impact of the COVID-19 pandemic to calibrate policy advice has made high frequency data not only more relevant, but often in the first instance the only way of measuring the impact of the crisis in real time. The swift economic policy responses was in part made possible by the existence of

42、 programmes to facilitate work-time sharing among employees and which in many countries were already in place and thus ready to be activated. Other programmes with features that could be made more contingent on the state of the economy include for instance unemployment benefits, or support to busine

43、sses in financial distress that would neither overload them with debt nor distort competition and which could help to limit the unnecessary liquidation of otherwise solvent and viable. The need to calibrate such policy actions on real-time assessments of economic activity increases reliance on high-

44、frequency indicators.The past few years have seen the emergence of new types of high-frequency indicators. These include flight departures, restaurant bookings, mobility reports based on anonymised personal data from Google and Apple, air quality indices, news-based indicators such as the Economic P

45、olicy Uncertainty Index (Baker, Bloom and Davis, 20132), electricity consumption, and credit card transactions. These new indicators are often available on a daily or real-time basis and for a range of countries. Policy institutions and National Statistical Agencies (NSOs) across the world have turn

46、ed to such alternative data, including the ECB (Benatti et al., 20203), the Bank of England (Bank of England, 20204), INSEE (INSEE, 20205), the Federal Reserve banks of Saint Louis (Kliesen, 20206) and Cleveland (Knotek and Zaman, 20207),and the IMF (Chen et al., 20208). Relatedly, the Harvard-based

47、 project “Opportunity Insights” gathered a large number of high-frequency data on the US economy from private companies. The OECD has leveraged a number of high-frequency indicators (OECD, 20209), including Google Mobility reports (based on the locations of Google Maps users). This paper focuses on

48、Google Trends data, which provides aggregate information from Google Search.3. Exploiting the full potential of Google TrendsIn the past decade, a growing literature has provided evidence of the usefulness of Google Trends data for nowcasting the current state of the economy (Varian and Choi, 200910

49、; Carrire-Swallow and Labb, 201011; Chen et al., 201512; Narita and Yin, 201813; Ferrara and Simoni, 201914). Papers have studied the link between Google Trends data and employment or unemployment (Baker and Fradkin, 201715; Fondeur and Karam, 201316; DAmuri et al., 201217), as well as consumption (

50、Morgavi, 202018), trade (Gonzales, Jaax and Mourougane, 202019), digitization (Pisu, Costa and Hwang, 202020), or housing prices (Askitas and Zimmermann, 200921; Wu and Brynjolfsson, 201522) and construction (Cournde, Ziemann and De Pace, 202023). More recently, Google Trends data have also been use

51、d to assess the impact of the COVID-19 crisis (Abay, Tafere and Woldemichael, 202024; Doerr and Gambacorta, 202025).Google Trends provides Search Volume Indices, which measure search intensity (number of searches for a given keyword divided by total searches) by location and period. Queries can be m

52、ade by keyword, category of keywords or topic. Queries based on keywords are language-specific and subject to ambiguity. Google Trends series for the keyword “apple” mixes up searches for the fruit and the company. Both categories and topics are harmonised across languages, and queries based on cate

53、gories or topics yield series comparable across countries. Google Trends thus provide a dataset of monthly panel data. Focusing on observations from 2004 to 2020 for 46 countries, the monthly data give a total of 8370 observations. Using topics and categories allows for more general models as topics

54、 and categories abstract away from keywords and provide a representation of search interest for things rather than specific terms.Google has classified searches into 1200 categories HYPERLINK l _bookmark6 2. These allocate individual searches to (multiple) categories using a probabilistic algorithm

55、(Varian and Choi, 200910). Categories are structured as a 5-level hierarchical classification. For instance, the category “Autos and Vehicles” aggregate together all searches related to cars, whereas an equivalent query based on keywords would have to explicitly combine each possible car name and br

56、and.Topics address the ambiguity problem of keywords. The topic “Apple (company)” allows to single out searches related to the company not the fruit, and combines searches for keywords such as “apple watch”, “ipad”, and “macbook”. Google has created topics that aggregate together multiple requests m

57、ade on Google Search based on their purpose and meaning, by taking into account where users click. There is no fixed list of topics and topics selection implied exploration from the Google Trends website.Categories and topics also have drawbacks. The exact content of topics and categories is opaque

58、and the algorithm that allocates keywords can make arbitrary choices. For instance, the topic “unemployment benefits” encompasses mostly French keywords in Canada, and is thus informative about labour markets in French-speaking Qubec rather than the whole country. This further warrants the use of ma

59、chine learning algorithm capable of flexibly capturing cross-country causal heterogeneity.The list of all categories can be found in this repo: HYPERLINK /pat310/google-trends-api/wiki/Google-Trends-Categories /pat310/google-trends-api/wiki/Google-Trends- HYPERLINK /pat310/google-trends-api/wiki/Goo

60、gle-Trends-Categories CategoriesThe present exercise exploits the full potential of the Google Trends data by using both category- based and topics-based searches. Two hundred fifteen categories were selected from 1 200 on a judgmental basis, as selecting data based on judgment may provide better re