Sign Idaho Banking Notice To Quit Mobile

Industry sign banking idaho notice to quit mobile

my name is jim naren i'm the first vice president and chief operating officer at the federal reserve bank of philadelphia so on behalf of everyone here at the bank i'm delighted to welcome you back to day two of this fourth annual uh conference so before we get started i just wanted to do a little bit of housekeeping uh so as i mentioned it's our fourth annual conference uh something new this year it's our first time doing this virtually so we really appreciate everybody uh hanging in as we all explore this new technology so that's something new this year uh what's not new is the standard fed disclaimer so the remarks that i give today are mine and mine alone they don't necessarily represent the views of the federal reserve bank of philadelphia or the federal reserve system so we've got that out of the way so i'm just really thrilled that we've got a stellar group of speakers today they're going to talk about the relationship between big tech and central banks and on the development of central bank digital currencies but before we turn it over to our panelists the the main considerations and themes that i suspect we're going to hear during our discussions this afternoon and my remarks are going to be brief uh now i've worked for four different reserve banks plus the board of governors the treasury and in the private sector and across a variety of of banking functions uh and what i can tell you from my own personal experience is that the fed and the private sector have a long very fruitful a cooperative relationship now much of that's based on sharing of information now we can learn from the private sector in particular big tech what's happening in the economy real in real time that's really important for us that strengthens our research capabilities and really bolsters our ability to formulate policy that best addresses current economic conditions now epidemiologists have long used search analytics to pinpoint disease outbreaks so if people in a certain part of the country are furiously googling flu remedies that's a pretty good bet that there's an outbreak of influenza in that area now likewise search analytics can inform us of localized economic events let's say a spike in search for unemployment claims in a certain part of the country or among a certain demographic group can tell us something important about what's happening in local economic conditions long before the official data would ever tell us so now one of our panelists that's going to be providing some pre-recorded uh remarks today hal varian uh chief economist google uh certainly is going to have something to say about this but this kind of relationship that i'm i'm talking about this relationship between big tech and the fed really predates this this kind of hyper sophisticated activity that i'm describing so back around the turn of the millennium back when google's employee head count probably numbered in the hundreds not the hundreds of thousands i was working at the new york fed running their cash operation when a tractor-trailer shipment full of nickels but our little five cent piece uh was robbed yes nickels so so imagine the disappointment with the thieves when they opened up the back of that trailer and found it was full of nickels now the feds had this long-standing relationship with a company called coinstar that's a company here in the u.s that owns these little green kiosks that people dump their coins in when they find them between the couch cushions uh so we asked coinstar and they have great data that to look out for any deposits of nickels the nickels come in two thousand dollar bag so they can see these increments coming in so we asked them to look out for any patterns that might be emerging they did they found just such a pattern in the southeast of the us where this robbery took place which eventually led law enforcement to the nickel robbing bandits so if you think about this this is really an early example of big data in action so big tech can help the fed gather information uh formulate policy uh even fight crime and as a general rule there's no doubt that increased access to data is a boon for policy makers like us but the relationship also raises important concerns about privacy a big tech covers up extraordinary amounts of data and information on all of us each day so tech firms and the fed are going to need to work to strike that that right balance between protecting people's privacy and using information in a way that serves the economy and our society now similar concerns are going to lurk with a central bank digital currency and that's the topic of today's second panel so first a quick note about how central bank digital currencies function so much of my time in the the fed has been spent running our cash businesses cash continues to be an important part of the financial system despite its off predicted demise it's still here cash demand has remained robust here in the us even during the pandemic but central banks and and the fed really issue uh two types of money both the physical currency that i'm talking about and reserves so think of reserves as the electronic unit of account that banks hold with a central bank so a new central bank digital currency would combine features of these two existing types of money so like cash a well-designed central bank tips digital currency would circulate widely in the economy uh it have used universal either low-cost or free access for households and businesses alike and like reserves central bank digital currency would exist in electronic form really making it easier than in physical currency to store or to use in in transactions across larger distances now there are obviously potential drawbacks as well the potential loss of privacy and anonymity now we all know somebody who prefers to use cash because he or she doesn't like their spending being tracked you know even if it's as innocent as that a daily latte at the coffee shop and what if someone's tracking my spending patterns to make credit decisions on me these are all issues for serious consideration and at some point potentially regulation and possibly legislation so as with big tech we'll need to strike that right balance that maintains the appropriate uh privacy protections while also promoting efficiency and and the benefits that we get now of course the federal reserve hasn't been a first mover in digital currency development for some pretty obvious reasons but it also appears that central bank digital currencies are coming uh one way or another which is why i'm delighted that we're hosting the conference today so thanks again to everyone for joining us and now let me turn things over to our first panel of the day which is going to look at big tech and central bank collaboration so we're extremely fortunate today to be joined by kern nemond anderson he's an advisor to senior management at the european central bank uh hal varian is going to be providing uh some prepared uh remarks via video he's senior economist at google and moderating our discussion the discussion today is going to be adam schabe of fna so thanks everybody for for joining us and and please enjoy the program this afternoon thank you very much james for a kind introduction it's a real honor to be here and many congratulations to julata jaktiani and her colleagues from the federal reserve bank of philadelphia for organizing an extremely interesting and extremely uh insightful conference ladies and gentlemen uh good day my name is adam chabai i'm the head of subtech at sna and it will be my pleasure to guide you through this session we are going to cover i think a very interesting topic as we are going to talk about the case study of a partnership between a big tech firm and a central bank and we will also discuss some of the key opportunities and implications for the mapping and monitoring of the financial system and of the economy i'm very pleased to say that we are joined by two seasoned practitioners who are very well equipped to guide us through this kind of collaboration our first speaker is pair neiman anderson adviser to senior management at the european central bank with over 20 years of central banking experience pere is also an author of several publications including a soon to be published book on data science in finance and economics our second speaker is halvarian chief economist at google and of course one of the key authorities in the fields of data science and economics before we start ladies and gentlemen i would like to encourage you to submit your questions through slido and we will do our best to cover as many of them as possible but now without further ado let us start with our discussion and i would like to invite to our virtual stage our first speaker neiman anderson from the european central bank welcome and many thanks for joining us at this conference thank you very much adam and thank you for this kind introduction i would also like to thank julapa yatiani and the federal reserve bank of philadelphia for this kind invitation including also the other members of the organization committee christopher ferenae richard henning aaron klein patricia moser and wagner van der roha rao for your kind invitation to speak about this the use of non-standard data and partnership and i have my i have my three messages i would like to convey to you first is that data never sleeps so there are so many alternative sources that are being used now and then i will make a little um a demonstration of how we have used the google data in order to now cast car sales and then the way ahead and as a central banker and i would like to say that the opinion that i'm expressing in this presentations are not necessarily those of the european central bank or those of any of the central banks of the european system so they're all on my own personal account so we have a beautiful time i mean the amount of fintechs platforms that exist and the digital transformation in finance and economics provides so with a huge paradigm of records that that are all digitalized and these can come from um digital trading so for instance the ecb and we are now producing daily yield curves of the aaa and eur and euro area [Music] bonds and they are coming from your mts in paris the electronic trading platform in paris and we are taking this data on a daily basis with their price information and then we are releasing the euro daily yield curves on the ecb website every day at 12 o'clock of course then we have the old clearing and settlement systems a lot of interest of consumer and credit card shoes mobile payments scanning data and what we will talk about today is the social media and the use of google data of course there are many financial platforms of alibaba and the various alipay and credit pay that exists trump's crowdfall funding where of course lenders and financial investors are meeting in order to allocate and distribute funds and then of course we have the big um blockchain diplo ledger technology crypto assets digital assets smart contracts tokenizations of various assets and token trading that are then also leads over to supply chain management and the ecosystem so there is enormous rich pool of digital data that that can be used and and this of course data never sleeps so this is just an illustration of how much digital data that is generated every single minute of the day and of course these non-standard high frequency data has now become more important they are of course triggered by the cuvette and and epidemic trigger to have more timely and daily data in particular to look at consumer spendings which you can use credit cards and debit cards also restaurant bookings can be an indicator air travel so they are different alternative sources and what i would like to demonstrate today is the the partnership that we have with google um using the google search terms of the exchange of knowledge that we have and from the central banking point of view we are a data driven organization um at least we believe so and we have a lot of experience in handling micro level data so we have a lot of micro data from all bank loans to other banks to the money market loans and we have securities issues and trades and we have bank loans to the corporate sector so we have a lot of microsoft we can obtain millions of micro data every every month and we are able to manage and systemize this data so we have something that we can offer then of course we have the analytical capacity and skills the central banks we are very much analytical driven and we are also familiar with the data science techniques and tools and and not last at least and for banks we are independent and therefore we provide object objectivity in the advices that we provide and then google can offer to us which they do it's all the google data all the google search data for 14 countries we obtain free they have also made a taxonomy of all the search terms so they're all categorized into different baskets and we are then sharing the insights of methodology and how we're using the data and such and we are automatizing machine to machine downloading of the data that we obtained from google so this is this is what would what i would like to demonstrate and one of them is of course that for central banking purposes we have an interest in looking at these alternative sources not only may be triggered by the kobe it has certainly raised the agenda item of it is to in order to get supplementary insights to get near real-time snapshots of the economical situation can help us with detecting early warning signs that may have and of course to help us to take trends and turning points and then of course the feedback loop from when we do policy measures what are the reactions of the the market and then of course the supplementary data could provide new insight that maybe we can adjust and maybe even develop new theories so better predictions stimulate the debate we all know we need to adjust our model based theories and maybe include some more cognitive processes in decision making and it's not so linear more timely and frequent information and this relates not only to central banking monetary policy arm but also to banking supervision as well so what we have done with the google data and this is the example we have been a category of the google data that we have used and we have taking all the search terms relating to cars and put them into a urary indicator called auto and verticals and then we have created a single indicator called then the google data and the google data we are then compared with the official car registrations by the car industry so this is the first thing you do of course you generate your indicator you you simply develop it and and then you simply plot it together with the official data from the car registration this is how it looks like so the red line is the google data sets the official one and it's the blue and then we simply see and look and then we start with a little bit econometrics and look at what are the dynamics of these two series so the first thing we see is that there is a certain co-movement of the of the two series in fact there is a two-way casual relationship so they both seem to in this example to influence each other and there seems to be a five lag length of of of the google data vis-a-vis the car registration so that seems like that every time we have here the google data it seems like that five years down the road there is a similar type of pattern so this is just by making some preliminary econometrics on the data to to get a feeling of your data and then we are then trying to do some uh some some now casting and here i have seven different uh ultra regressive models that we have used i will not go into the details but they are relatively simple we have a basin baseline model that is simply taking the car sales and then using the seasonal um data from 12 months beforehand and then we see how good it is as a predictor and then we include it with the google data and then we try with other types of indicators we use the cpi so the harmonized index consumer products we're using industrial indicators and the postal income of households and plus and my plus the the google data so how do they then then perform and this is the results um o all the the tests that we have done so we have all we have the baselight model on one side and then we have all the different models that we have tested baseline with the google then inflation rate with the google and then confidence indicator with the google's income with the googles and so on and then we have just tested all indicators and then across we have taken criterias as of relating to the error terms that we have which are the the the root mean square rows the mean absolute arrow and the mean absolute percentage error which is the forecasting area and then we simply take an average of how much more or less the improvements compared to the base model have been so there's five takeaways from from this little exercise we've done one thing is that just using the baseline model as we have seen seems to be quite well it performs quite well just using a baselight model the second thing we see that every time we add the google data the improvement seems to improve so including the google data in our forecasting improve the error term so it reduces the the arrows the third is using the model with income and google seems to reduce the error terms by 30 so that is the highest so the best performing model and of course if we then take all the integrators together there's a slight increase and then with the google data including we simply reduce the the error term significantly the fifth is that when we use the default and mario test so that is to say that if we compare the forecasting accuracy using this data set can we statistically say that there are significantly better predictors using the google data then not using the google data and for this everything that is highlighted in red indicates that it's statistically significantly better predictors than not using the google data and that seems to be the case for every single time we have used the google data that they all statistically statistically significantly perform better than the base like model and the sixth takeaway from this exercise is that we do exactly the same test so exactly the same depot the mario test pairwise so i'm looking just as the model inflation rate inflation with google so pairwise and there i also see that including the google data in this model compared to the model without the google data the model also performs statistically better than without the google data and then there is a little outlier one could say that for the household savings and the sales of saving with the google data i need to go up to the 20 acceptance significance level before i can then reject the h0 hypothesis is that the data sets or the data models are the same so these were a little taster of the data set we have with google here we can then see the results of the google baseline model and then the best model so the model that performed the best so that is the one in red and then the base moon is yellow and the official data is the one in blue so having this experience and maybe of course also other experience i just wanted to make one little note that i need all the time and i will take the the the time here to simply say that just because we have all the google search term data or just because you have all the data from a certain fintech does not necessarily mean that we have to circumvent what is sample bias and representativeness and that seems to be a misperception that is very often read in also research books research articles and different science that we simply don't look and engage ourselves in the data so for instance the google data it's a recording right so it's people search but they're more there are the search machines right so if we're just using the google search data and we also have to see who else are using other types of search engines so they may not be fully representative of all the search term though of course they seem to have a certain monopoly or a large share of the search engine now of course it's not only households that searches right so there's also corporate uses so if you're looking at household expenditure or households expenditure in this case for for the car sales there are also a lot of searches that doesn't relate to households and then of course we still have certain areas where the internet connections are not very good and not all types of people are using the internet and maybe there's an age and other types of criteria of of people that we we don't get and then of course the second issue is that we are not measuring people right so it it's a it's a unit measure it it's it's a search it's a search term that means that one person can make 20 or 30 searches so therefore it's it's not about people but it's about searches and then of course using uh search engines it's event driven right so there can be many searches that are happening in a certain period of time we have for instance the the card mission scandal and then everyone is searching if their car is impacted by the diesel or if they have this motor or this car and so they're not driven necessarily by the people would like to buy a car but you have to then of course remove the data that that will in this case be considered as outlier so my my three takeaways to you from this presentation is that we have to experiment with these alternative sources and these experiments they they do not bring short-term results but we will in the long-term uh see the benefits of these and they are valuable sources both for economic and financial activities and we need to think about digital data strategy and how to make data more borderless we have in the eu european data strategy we have to find out not only about the data but it's also about the metadata and how to share data cross-border while of course preserving the the confidentiality of data and it's about moving from experimenting with the data to have it part of a toolkit that we can then use it as part of giving advice and certainly my encouragement is to leverage between the partnership between fintech and central banks but also the academia as part of leading for excellence so thank you very much thank you very much per for your very insightful and interesting presentation uh i have a number of questions i would like to ask but in the interest of time i will leave them for later and now ladies and gentlemen we are going to focus on the contribution of our second speaker halbarian halvarian was unfortunately not able to join us in person but he kindly agreed to pre-record his remarks and we are going to watch the pre-recording of his presentation as well as of the discussion and the q a now thank you very much for inviting me to this meeting i'm sorry i couldn't uh attend in real time but i hope the video will be some consolation so i went to talk today about google tools for now casting and uh let me start with a little example uh we're going to look at data from google trends and google consumer survey and then we'll look at some analytics and tools that we've developed here that are available to you for your own analyses and i start off with a question for everyone which day of the week are there the most searches for hangover and how would you find out well you go to google trends type in the term hangover and you see that actually they peak every sunday and what's that great big giant spike that's actually new year's day uh you can look at it geographically so here's the incidence of queries for hangover across states in the u.s and you can see the kinds of terms where people are entering in for this particular case the hangover cure hangover hangover energies remedies and so on and here's a plot of the hangover series in uh some color here red i guess and queries on vodka so uh if you look at this off just by one day the vodka searches peak on saturday and the hangovers on sunday this is a beautiful example granger causality if you're into a time series a little picture there churches for civil war why do they have that particular shape well it's seasonal turns out the peak here is about three days before your term paper is due and the reason the series has been going down is we're measuring the searches for civil war those particular queries over time and during this time period the internet went from being mostly something that was used on campuses is something that's used everywhere so it isn't a decline in absolute interest it's a decline in relative interest in this particular uh example uh there's another little picture of what searches return paper look like and this is a nice example looking at uh gifts for boyfriend in blue and gifts for girlfriend in red and the little drop there is christmas so you see the um activity among the men who are searching for a gift for girlfriend that explodes right before christmas because they see they've got only a few days left to get a gift same thing get for husband gift for wife looks like this and so on so there's a lot of fun things you can do with google trends of looking at what kind of searches people are conducting but there are also some serious things where you can get some insight into economic activity so you can do time series forecasting you can do cross-sectional prediction where we look across cities or across counties so on and you can do various kinds of forecasting or now casting that i'll describe in a few minutes so forecasting that's predicting the future now casting that's predicting the present and why would you want that well as you all know economic data is released periodically on a monthly or sometimes quarterly basis data revised as new information becomes available and what now casting gives you is a way to look at what's happening right now in the short term and that can be very useful because the data can be more timely and certainly for cases involving central banks that's uh that's something that's extremely attractive so here's an example this is initial claims for unemployment in black and in red you see the unemployment rate the headline rate and you can see that the initial claims is something of a leading indicator for the unemployment rate and when you look at people who do forecasting one of the most important numbers to look at is in fact the initial claims for unemployment uh because it gives you a high frequency view of the of the economy now what would you do oh here this is a picture of the relationship time series of initial claims in blue and the unemployment office unemployment filing in red and you can see the queries on unemployment filing in fact seem to be very highly correlated with the series for the initial claims and that's pretty reasonable because if you became unemployed the first thing you might do is go to google and type in unemployment office how do i apply for unemployment what do i need to apply for unemployment where's office located et cetera et cetera and those queries are of course expressions of intent and the intent is to file for unemployment which is what's uh what's going on with this particular time series so one thing you could do is you could predict the initial claims for unemployment from fred or whatever uh using lag values of initial claims and contemporaneous queries on unemployment filing so you could first run a very simple seasonal auto regressive model that's what i call the base and then you can throw in to that the google trends data the query data on uh unemployment filing and it turns out the r squared goes up the aic goes up definitely there's predictive power in adding that uh initial claims queries that is the searches uh to your uh model so we discovered this i don't know 10 or 15 years ago a long time ago and we built some tools that help you carry out this exercise so for the particular one is something called bsts which is bayesian structural time series where we use a common filter for modeling the timing component the seasonality and the lag variables and the structure from the time series and then we use a regression uh component what's called a spike and slab regression for looking at the role of the google query from the trends data so i'm not going to go into the econometric details i just wanted to mention those two things because anybody who wants to follow up can look at the literature in this area you can get it from ssrn or my website at berkeley wherever so the idea there is you have a trend plus a seasonal plus a regression and what you do is that's your model as a as an econometrician the uh regression is where you put in the data from the google trends uh the trend and seasonal or your usual time series uh modeling so you look at seasonally adjusted data or non-adjusted depending on what your preference is this is a very uh packed slide on uh coleman forecasting i'm not going to go into it in detail but just point out that it's a also gone known as state space modeling and it's a very handy uh tool particularly for this particular problem uh again i'm going to flip that as skip this slide because it's uh too detailed for the amount of time i have this is what the code looks like you say why is my data this response variable might be the initial claims and then you say add a local linear trend so i'm going to put some smoothing in there add a seasonal i'm going to put in a daily uh day of week effects seven day seasonal adjustment i put that together pull that together as my model and then i click the button and off i go now the question is does the google data help explain those initial claims and if you look at that picture up at the top we're looking at what happens if you just estimate using time series no trans variable that's in black and with the trans variable it's in red and you can see there's a definite reduction in the error rate particularly particularly around the onset of the great recession and then you can go back to that uh model that framework and look at how much of that change was due to just trend how much was due to seasonal and how much was due to the regression that is how much was due to the trends variable you can see the regression really explained a lot some of those three series are the actual uh well the actual um yeah the sum of those three series depicted as the actual and you can see the regression is doing playing a big role so that's very helpful in describing the actual outcome variable and we can see which variables were important so what you do here is you throw in variables you think might be important and then it reports back to you on the probability that they are in the regression so here if you look at it unemployment office is the uh query that's the most helpful in predicting these initial claims filing for unemployment idaho unemployment then a couple sort of spurious regressors down there on a serious internet radio these are not don't have a causal interpretation of course uh the idaho is not particularly important either because there are queries issued for the unemployment offices in every state and that one just happened to have the highest correlation you do the same thing with other data sets this is new home sales in the us you've got the 80 20 mortgage is the google trends variable we chose and these are the non-seasonally adjusted housing numbers again that i pulled down from fred here the variables that are predictive are queries on appreciation rate irs 1031 century 21 realtors real estate purchases and so on so in fact the real estate related queries are important in helping to forecast the or now cast of real estate activity uh i'm going to zip through this pretty quickly i just wanted to show how the model improves as you add these additional explanatory variables so here we've got just the raw data and some fitting a simple trend to it no predictors at all here we add in the seasonal component so you can see how that varies then we add in appreciation rate the irs 1031 the century 21 realtors real estate purchase and you can see you've got a very very good fit now this is in fact an over fit uh you don't want to um it doesn't necessarily have predictive accuracy but let's see if we can fix that a little bit uh as through an 80 20 mortgage here uh do a one month ahead forecast so i'm now predicting out of sample not looking at the in sample prediction and it does 23 better than a simple autoregressive model another thing you can do is ou can use query categories instead of using specific words you can use uh categories such as real estate agents home insurance etc etc and there we have a picture in the top of looking at how your model improves when you add in those commercial queries and you can see the raw series down below there's a nice feature in bsts where you can fit the model up until some point and then freeze the model and just go ahead and forecast out of sample using the same model that's what we did in this top graph so the model was estimated on the observations to the left of that vertical bar and then the model was frozen no more learning no more estimation it would just use the the predicted model that was finalized let's say at that at that point so that's an example again of out-of-sample prediction it worked really pretty pretty well here's the same thing for motor vehicle sales same idea here i use categories so autos and vehicles vehicle shopping that comes with the trends data every query is categorized and you can see the same general idea works there build the model freeze the build at some point forecast the next however long you want to go and you can get a model that does now casting and in this case some forecasting as well let me give you a geo example this is a not well sort of an economic example uh there was some nice work by ed glaser at harvard and his colleagues on the happiest cities in the us and uh the cdc has a behavioral risk factor surveillance system that asks people how satisfied are you with your life and you can look at the responses by city so you might want to know what queries are predictive of happy and unhappy cities so we ran a little geographic example and here are the predictors gambling that's a big negative predictor and it's not casino gambling or anything glamorous it's basically lotto uh if there's a high frequency of queries on gambling in a given area you could be pretty sure people aren't going to be very happy in that area manufacturing insurance etc so you can the black lines i should have explained this so black lines are the negative uh predictors and the white lines are the positive predictors so there's your regressions for happy cities every time i give this talk people say well which are the happiest cities in the u.s uh here's the actual and predicted based on that model and the happiest city is charlottesville and the most unhappy city is new york maybe you already knew that i know it was a bit of a surprise to me um another tool i'm going to leave google trends for a moment uh is google consumer surveys and i'm sure you've encountered something like this as you've looked around the web to be a survey that says please answer this question to proceed something of that sort there's a very short very simple one question survey people answer that survey and then they go on uh doing whatever they were doing well those surveys are posted by google and they are a fantastic way to do a very quick survey of the population uh lasting a couple days getting a few thousand responses that can be very helpful in judging things like consumer confidence and other metrics of that sort anyone can do these queries the cost is dramatically lower than conventional conventional surveys the results come back in just a few hours if you don't believe somebody else's survey you could replicate it which is very rarely done in the survey world but now it's possible and you can look at things where you're designing a big survey and you want to look at sensitivity or questions around wording or framing for some specific components of that series so here's a question from pew who asked if you were asked to use one of these commonly used names for social classes which would you belong in and you see a response of that sort but 45 percent of the population says they're middle class 21 lower middle class and so on we asked the same question and we got phenomenal alignment with the google consumer surveys and the pew forecasts so this is responses from google and the responses from few and if you go to the pew website they do an analysis of the consumer surveys with respect to statistical merits and it's been very uh it does reach quite well in terms of performance here's a question on assembled in america i prefer to buy products that are assembled in america strongly agree 32 percent agree 30 and so on but you also get all the geographic data from this okay that's the really nice thing that you can look at it by geography you can look at predictors of survey responses in regions in places where a lot of queries on chevrolet firearms and weapons country music and trucks people strongly believe in the somewhat in america probably not too uh probably not too surprising and these are the top cities on the left which are basically south carolina virginia north carolina and the ones on the bottom which are basically all the west coast uh california and and seattle and so on there's a picture of the places where people strongly believe and assembled in america again probably alliance with what your expectations are but there's a nice way to quantify and visualize it um you can do the same thing for any geographically distributed query i'm not going to go into this either this is hard places from the new york times index and we could look at what kinds of queries are associated with hard places what kind of queries are associated with easy places and you can do all sorts of stuff this is a nice little node from some researchers at the board of governors on predicting recessions at different forecast horizons it uses a similar model to what we use not the same thing but similar and finds out what predictors seem to be the most helpful in predicting the session onset at different horizons so i'll leave that one for you to follow up and that's the end so thank you for your attention it's been a pleasure speaking with you great thank you very much hal for your extremely interesting and insightful remarks i'm sure you have provided uh our viewers and listeners with some thought-provoking takeaways uh i myself have like a dozen questions i would like to ask and i'm sure it's the same for pair but in the interest of time uh let us keep it to two or three questions each my first question hal touches upon collaboration uh between your team and google and various central banks and i would like to know uh what do you see as the key value or contribution of this partnership to the economy or the financial system and if there are any successes or achievement that you would like to highlight in this regard well one of the things you can do is you can go to uh google scholar and type in this now casting or bsts or any any of the terms that i use in the discussion and you'll see dozens if not hundreds of papers that people have worked on at central banks and at other places so for example [Music] one of the things that you can do with this system is it's very very good at predicting um vacation activity particularly in europe so people in march and april are planning their vacations and then they take the vacations in july and august and vacations can be very very important for some countries and some cities for example in egypt is 12 percent of gdp and this is something that the bank of spain has worked on bank of israel has worked on several of the uh central banks in europe have built models for vacation planning uh vacation forecasting and it's a case where it's not just now casting it's true forecasting because you're looking at these searches now the intent that's going on now and the outcome is going to be a couple months from now great thank you very much al my second question is a bit more forward-looking this conference is being attended by hundreds of central bankers from various jurisdictions and i'm sure many of them would be very much interested in hearing your views on which areas of technological innovation you would suggest for the central banking community to pay special attention to in 2021 well personally i think it's very interesting to look at regional macroeconomics that is getting data at a regional level and thinking of a place like ohio as being a small country that trades with its neighbors and and has this kind of regional dynamics that we also see in uh larger country level analyses so i think having fine-grained geographic data about economic activity is going to be very helpful in improving our understanding of macroeconomics great thank you very much hal in the interest of time i will stop here but i'm sure you also have a number of questions that you would like to discuss with how so please over to you thank you very much adam and how good to see you again and uh i have two two questions for you if i may one is as you may know that in europe we have now launched a new eu open data strategy for creating one market for sharing digital data what is your view in this initiative and what are the enablers as the next step for digital's data to become a public good very good question and we've been working on related issues ourselves about four or five months ago we released something called data set search which allows you to go to google type in data set search you'll be taken to a search box where you can look for data on various uh topics so i would strongly encourage you to use some of the metadata that comes out of the data set search project because it will make this data much much easier to find and i i think it's an open standard everybody can use it it's just a way of describing data sets in a way that they can be easily found also just as a as a footnote on that natasha noy that's nly is the researcher of berkeley who i sorry google that built this and she has a very nice little paper full of descriptive statistics about how data is available on the internet so that would be worth looking at as well right oh but that i i certainly can see this l that the metadata describing that content is of course vital for being enabling the sharing my my second question um hell is relating to um private public partnership so maybe our collaboration that we have between the ecb and google may be a good example of this but what are the enablers according to you your view for increasing the data sharing between academia and the private and public authorities yeah so um as you know the imf and the world bank just had their annual meeting maybe 10 days ago something like that i spoke at that meeting as well and there's a great deal of interest in this in this question i'll tell you our view at google or maybe my viewers google is what we like to do is give the data to everybody or to nobody because those in between cases if you have to make a custom provision for one party or another party then it gets to be quite uh consuming of resources and uh it's much better to have a uniform policy that people could use to access this data so anybody could use trends anybody can download the csv file anybody can embed the maps and so on but if you're a non-profit a government agency research center then you can also get access to the api which allows for more flexible way to query the database and i think i think the best thing to do is to start small try a few small projects see how they work gain some expertise and understanding and then branch out from there because i'm sure these relationships will evolve over time as people learn the value of this kind of data thank you help so uh so when we uh related to that uh pair uh and how i thought we would also uh like to hear your thoughts on um data sharing but there have been concerns about consumer protection and all that what would be what's your thought on the right regulations to on data sharing that would protect consumer privacy but at the same time avoid discouraging printed innovation right so in this case we spent a lot of time roughly months six months of time figuring out a framework for releasing this data in a way that protected privacy so to show up in the database at all you have to have something like 40 or 50 instances of the query from separate ip addresses so you can look at situations where it's a relatively common query now this has pluses and minuses it's a plus because it's almost impossible i mean it is impossible to access any kind of individual data because of that aggregation choice but it's also problematic in the sense that for some kinds of research it is important to be able to get down to find granularity like for example epidemiologists epidemiologists want to look at the progression of disease and that's become particularly important these days because of cloven 19 but um what happens is they can't follow it down to its source because you're getting fewer and fewer people issuing queries as you go down so we've kind of stuck with looking at things at the county level uh or at sometimes at the metro level where's places where aggregation is uh is your friend and it's a not possible really to reverse engineer any uh individual data great thank you very much for your response and uh i would be behaving this is awesome great i'm glad it worked out i have to say i was a little uh skeptical but it worked very smoothly so maybe i'll give another talk next week great so many thanks again to hal for his pre-recorded remarks i'm sure he has provided a number of inspirational and thought-provoking points uh i'm now looking at the watch and i think we do have time for two or three questions before the session concludes uh pair there is a delegate question which relates to the area of consumer protection that we have already briefly touched upon and uh aras would like to know what regulations we have or should have in place to make sure that the big tech firms do not misuse their access to the private big data thank you very much for that question i think this this is vital this is a quite a fundamental question and we have the general data protection regulation in place so of course when we are talking about borderless sharing or simply sharing of data that has to be ensured that it follows the general data protection regulation in addition to this of course in particular central banks and other public institutions are of course used to protecting confidentiality so this is our bread and butter so of course we have a lot of confidential data and we have certain methods in order to protect and there are also data protection confidentiality regulations in place and of course when we talk about sharing data these has to be fulfilled so this is not only related to consumers but certainly also to corporate so any way of one can identifying individuals or individual firms has to be removed from the sample if they are too dominant or if it's feasible also with combining other sources to identify the person so this is absolutely fundamental and it's a precondition now having said this there are various mac regulation methods that is available and can be used as part of aggregating the data to a very low level while still preserving the statistical properties of the data sets that it can then be used for instance for research purposes and then of course there's also various methods for anonymizing individual data but of course the general data protection regulation shall not shall not be in our way to have a borderless strategy for digital data simply because the volume information is is so large that that we would need to ensure that data is feasible for it to be shared among different countries and there of course there the the european commission has come out with a white paper or having a having how to share data within the european union for instance and there are certain conditions that has to be fulfilled also for sharing them outside of the european union so there is a good start of this but i think the question is so relevant and i think every time we speak about this we should always say that it should never circumvent the general data protection regulation and the protection of confidentiality of individuals before we start using the data or sharing it with anyone else so thank you for that good question great thank you very much for your response and many thanks again to arash for you for the excellent question uh another question i would like to ask there relates to the disruptive era o the covet and of the covet crisis and the post-covered world that we will all need to deal with and maybe could you provide some examples of how the collaboration or the partnership between the big techs and the central banks can provide added value or benefits to the economy and the financial system okay yes yes thank you adam and thank you for that very valid question of course the corbett has simply been a trigger point to try to get more timely information of the impact in particularly of consumer spending so where how can you track consumer spending so for instance i mean we we actually know in europe people are saving so you can see saving rates going up so people are not really spending they're saving uh as such but if you would like to track it of course you could use a credit card data and payment systems mobile pay supermarket scanning data and if you have a representative a large market you can also look at the lending and the financial platform and and see who is borrowing and who is who is lending from these platforms so the market is segmented also according to the digital platforms um that amazons of these worlds right um there you can see you can use this data to see which category of products are being sold and how much and what are the differences if you had access um other examples is cargo transportation right so the postal services parcels and boxes how much are they moving around and and how much is the value of these so there are ample opportunities um only of course if the data is standardized and structured in a certain way that it can it can be read and produce time series and of course in addition that we have to ensure that we have the protection of confidentiality so no individual person can be identified or being identified by combining with other other sources thank you adam great thank you very much pair for his response as well your remarks brought that pretty much to the end of the session so thank you very much again for your very interesting and insightful presentation and also for answering uh the questions from delegates and myself it was great to have you here thank you very much and thank you for the invitation uh also i would like to say a big thank you to hal varian for pre-recording the presentation and to his colleague barbara phillips for making all this possible and last but by no means i would like to thank you ladies and gentlemen for your attention and for your questions and contributions i hope you have enjoyed the session and i would like to wish you a great rest of the conference and a very nice rest of your days thank you very much all the best and goodbye