 So welcome back, everyone, to this second session again on Microprudential Policy Effectiveness. So I'm very happy to have here with me the first speaker, that's Celine Deschreider. She's assistant professor at Kent University and has research interests in monetary policy and macroprudential policies with a focus on borough-based measures, as I understand. And her paper will be discussed by Sona Bashkaya, who's lecturer at the University of Glasgow and has lots of experience at the central bank of Turkey in various functions. So, Celine, the floor is yours. OK, good day to everyone. I'm very happy to be able to present this joint work with Lara Kulir, who's also present in the audience today, just in front there. So this paper is about trying to evaluate the effects of national macroprudential policies related to the housing market on house price growth rate. So more specifically, we want to find out whether there are heterogeneous effects of national macroprudential policy actions targeting the housing market, which are being implemented by the National Bank of Belgium on house price growth rate at the local level. So that's the idea of this paper. So let me then continue with providing you our research focus. So the question we would like to answer is the following, whether and to which extent those housing sector-specific macroprudential policy tools have heterogeneous effects on these local house price growth rate. And if we find them, then which factors could drive that heterogeneity? It's not only about finding heterogeneous effects, but also trying to see what drives these effects. And we put a focus on heterogeneity, or potential heterogeneity, being driven by local housing market characteristics and household financing conditions in those local housing markets. So the motivation for this research question is based on the related literature, where we typically see mixed findings on the effects of these tools on house prices, in the aggregate, meaning that typically people have been finding negative, small effects or insignificant effects. But it might be that those aggregate impacts are blurred because of heterogeneity, because we know from the literature that housing markets are segmented in nature. And they also respond differently, or they are differently sensitive to common shocks. And apart from that, there is also a policy relevance in the fact that whenever we find heterogeneity of these implementations on local housing markets, that this might have distributional effects, which is, of course, very important for politics, but also for policymakers, more generally. So these housing sector-related microbial tools are being used to dampen the so-called vicious feedback loops between the housing market, financial intermediaries, and the real economy. And they take different forms, but essentially they want to improve the resilience of financial institutions to housing market volatility. And they take different forms in the sense that it can be capital requirements that try to have a direct impact on this resilience, but also these borrower-based tools, which more or less have an indirect impact on the resilience of financial institutions, where we have the LTV caps that try to put an entry barrier in the housing market in order to reduce the share of risky borrowers in the mortgage market. But we also have the debt to income, loan to income, debt service to income limits, which try to reduce the vulnerability of households throughout the mortgage. And by reducing this vulnerability of households to negative shocks affecting the economy, also the resilience of financial institutions can be improved. What we are aiming for, however, is not this impact on the resilience of financial institutions, but the impact on house prices. Because typically these measures have been popular in the sense that monetary policymakers have implemented them relatively often. But at the same time, the main focus in terms of effectiveness of these tools has been on this resilience of financial institutions, whereas we know that also house prices can be affected. Because these tools have an impact on the financial conditions of households and financial conditions are the most important driver of house prices. So we in this paper really want to focus on the effects on house price growth rates. And so over and above this effect on the resilience of financial institutions. So we want to make two points first for the reason of potential heterogeneity on the effects of house prices. First is that housing markets are known to have a local nature. House prices are driven by many demand and supply factors. And typically these factors vary at the level of the local housing markets. But also we want to take into account that macro-prudential policy does not merely target the overall mortgage market, but specifically the high risk segment of that housing market. So also there, there could be heterogeneity driven by this specific segment of the housing market. So what we do is that we will use, in order to determine these effects on local house prices, we will use municipality level data from Belgium. And this for the sample period running from 2012 till 2020. So before I continue with an explanation of how we measure the main variables, being macro-prudential policy and house prices, I want to first give some further information on the Belgian experience in terms of these housing markets related policy tools over the timeframe we consider. And basically the Belgian implementations concern three different actions in this time period. The first one in December 2013, where there was a capital requirement, they need a fixed or flat 5 percentage point add-on to the risk-waste for banks using internal rate-based model. So this was just a capital requirement imposed on these financial institutions. And the second one was introduced in April 2018, where there was actually a strengthening of this risk-quade requirement, where next to the targeted component, which is again this 5% add-on, we also now had a risk-sensitive component, which is based on the micro-prudential risk-quades of the portfolio of residential relastics of Belgium, and then multiplied by a factor of 1.33. So the aim at this stage or at that stage was to further stimulate banks to reduce the extent of high-risk borrowers, the high-risk share in the mortgage market, because that was considered to remain too high despite that initial risk-quades measure. And then the third one was introduced, are effective since January 2020, and that are the borrower-based tools, namely prudential expectations on banks and insurance companies, which imposes a cap on the LTV ratio of 90% for owner-occupied housing and 80% of buy-to-let loans, albeit sometimes in combination with a debt-to-income and debt-service-to-income ratios, where then we have the cap on the LTV ratio of 90% combined with DTI ratio of 9% and then DSTI ratio of 50%. So basically three different macro-prudential actions oriented specifically at risk arising from the Belgian residential relasted markets. If we take a look at average house price growth rates during this period, we see that similar to many other European countries, these have been increasing over time. So what you can see on this figure is growth rates for the three different regions of Belgium, being Flanders, Volunia and Brussels, and then the weighted sum for the entire country Belgium in blue. And basically what you see is this rising trend over time in average house price growth rates, which has some variation across the regions, but yeah, the evolution is fairly similar. Then I want to come to the measurement of house prices on the one hand and macro-prudential policy tools on the other hand. And the most important thing to notice about measuring house price growth rates for our paper is that we use a hedonic house price index. We use data from a work of Ryssens and Quater. And yeah, there are many technical details associated to the use of a hedonic house price index, but the most important reason to use it here is that it allows to take into account the fact that the type of properties that is being sold in a certain period that that can change quite drastically over time. And this hedonic house price index allows to pick up changes in the type of properties that are being sold over time because we specifically control for characteristics of the dwellings. So it allows us to come to a measure of price changes for the average dwelling in the economy of Belgium. So this means that it has the benefits of leading to much lower volatility compared to a house price index which is based on transaction prices. And that's a huge benefit, but at the same time, it also has a drawback because we move from cartily frequency with transaction prices to animal frequency with a hedonic house price index and because of, so that leads to less observations. And also because of the fact that we want to use this hedonic house price index, we have to remove certain transactions because not all information about these housing characteristics is available for each single transaction in the Belgian economy. So whereas we have a measure which has much less volatility compared to house price index based on transaction prices, we have to move to animal frequency and we have a lower number of observations due to missing data on housing characteristics. On the other hand, we still find that if we look at this house price index based on hedonic prices, that even in case or even that's taking into account that even with this hedonic house price index, there is a reduction of volatility. Even in that case, we see that for small municipalities, there's still quite a lot of volatility in this index and this is being driven by a low number of transactions. So for that reason, we clean this house price index further by removing those municipalities from the sample which have a number of housing transaction in a given over the entire period which is below the 10 percentile for the entire group of municipalities. So we arrive to a measure of house price evolutions which takes into account the composition of houses being sold. But on the other hand, it's also very nice to interpret. And here I want to show the, I want to visualize the extent of cross-sectional variation because we use hedonic house prices to have a good sense of how house prices move over time, taking into account the composition of houses being sold. But we also noticed in our data sets that for the 497 municipalities we are having in our sample that there's a lot of cross-sectional variation in terms of the ranking of the municipalities over time. So it's not just that we have variation over time in house price growth rates at the local level, but we also noticed that there's cross-sectional variation changing over time. And these figures visualize this cross-sectional variation over time by providing you a picture of the map of Belgium with the different municipalities for the three, for three years in our sample being the first year 2012, the last year 2020 and the middle year 2016. And what you can notice here is that, so the color scheme represents quantiles of the distributions of house price growth rate in a given year. So we move from red to dark green, where red signals the lowest quantile of house price growth rate and dark green is the highest quantile of house price growth rates in that specific year. And we notice that there is substantial variation and typically, Mia and Lara give the example of Knokke, which is the most eastern country at the coast there, which is red in the first year, 2012, with a growth rate of minus 3%, turns orange in the middle year, 2016, with a growth rate of 0.5%, and then pops up to be in the dark green area in the last year with a growth rate of 17%. Of course, that's a pandemic year, we're talking about a municipality at the coast, so a lot of demand for second homes, of course. In terms of measuring my credential policy, we make use of a my credential policy index, and this index is based on the typical dummy being used in the literature, where we have values of minus one, zero and plus one, where minus one refers to loosening actions, zero to neutral actions, and plus one to tightening actions. So this dummy provides a very clean, elegant way to inform us about the timing and the direction of the change of my credential policies, but what it doesn't do is giving an indication about the intensity of those my credential policy decisions. And therefore, we use an intensity-based image of a my credential policy, which is based on earlier work of me and Lara. So I don't have the time to go into the details here, but basically we take into account the scope of each policy action, the quantitative restriction of each policy action, and the legal enforceability to be informed about the exact intensity of the measure over and above the timing and the direction of the change. Okay, and this leads to the following figure where we present the cumulative changes. So the risk wage in 2013, the increases in 2018 and 2020 associated with the borrower-based tools. And here I normally also want to show the additional graph to say that in the econometric analysis, we take the changes of these macro-dential policy measures, which are, and we take them together in one way. So we take together the risk-rate measures and the borrower-based tools in one series in order to pick up macro-dential policy changes targeting the housing markets in our sample. So having explained the two basic series and how we measure them, so house prices and macro-dential policy, I now want to outline our baseline model estimation. So what we have here is a dynamic fix-effects model where our dependent variable, YI-RT, represents local house price growth rates. So house price growth rate in municipality I, region R, at year T. We regress that on its lag, a measure of explanatory variables at a local level and then an interaction terms that picks up this macro-predential policy indicator where that's kind of the tool, this beta coefficient where we want to measure those potential heterogeneous effects of macro-predential policy affecting local house price growth rates. We take into account municipality level fix-effects, time fix-effects and region time times fix-effects to pick up any other confounding factors. So to repeat, we have almost 500 Belgian municipalities in our sample and we have a timeframe running from 2012 to 2020. So we use a dynamic or we estimate a dynamic panel with fix-effects. This means that we have to take into account the nickel bias on the coefficient of the lag-dependent variable and we do that via a jackknife bias correction approach which is shown to alleviate the concerns even in small t-samples. So first, it's X variables. So what we want to do with it is controlling for local drivers of house price growth rates being three measures of demands at the local level and one supply measure. So the amount sites measures are changes in per capita income, changes in employment and changes in the number of households relative to the number of total residents to pick up traffic factors changing over time. And as a supply determinant, we use the number of houses relative to the number of residents and exactly the pro-trade of that variable. Our interest lies, of course, in the coefficients on the interaction terms. So we interact this macro-prudential policy changes with indicators of financially constrained and high-risk borrowers, with a measure of activity in the local housing market and the degree of household indebtedness also at the local level, so at the municipal level. So let's explain this in more detail. So first, we interact macro-prudential policy changes with this measure of high-risk borrowers, residents in these local housing markets and we focus on different variables based on literature. We focus on the share of low-income inhabitants, the share of first-time buyers, the share of overdue credit relative to total mortgage credit, outstanding credit to young people, share of single-person households and single-parent households. So that are all indicators that signal high-risk borrowers are increased financial constraints. We also include the share of highly-educated young people and this is then a measure of residents being less constrained because it are typically those borrower at the age of buying a first house which are less constrained because of their relatively high education. Important to note is that we consider these variables to vary across these local housing markets, so across these municipalities, but they are taken to be constant in time and they are predetermined and this to limit reverse causality, namely the fact that those factors could also drive house price growth rates. The second type of variable that we use in these interaction terms with macro-idential policy changes is this hotness indicator where we take into account the number or the growth in the number of housing transactions relative to the number of residents in that local housing market and the idea here is that we want to see whether the hotness of a local housing market matters for the impact of macro-idential policy changes on local house price growth rate. And then the third measure of these interaction terms is household indebtedness at the local level where we want to pick up whether higher depths of households leading to higher risky position in that local housing market has an additional impact on the effectiveness of macro-idential policy again on local house price growth rates. So the hypothesis that we make here is that if you have local housing markets with more financially constrained, high-risk residents or borrowers that that will lead to an additional downward effect of macro-idential policy changes of a macro-idential policy tightening. Secondly, we conjecture consistent with findings in the literature that whenever you have a hot local housing market that that would lead to an additional downward impact of macro-idential policy changes on local house prices. And third, we hypothesize that when you have increased household indebtedness at the local level that again you have a larger downward effect of macro-idential policy tightening on these house price growth rates. So let's take a look at our findings our empirical estimates. So the table that I show you here is giving you the coefficients when we interact each variable separately. And from column three to column eight onwards you notice that all indicators of financially constrained borrowers, residents that those have a negative impact. So an additional negative impact of tightening macro-idential policy. They are all negative and except for the share of single parent households they are also significant. If we look at the first two columns however we do not find significant coefficients. So the first column shows you the result of trying to find out whether a hot housing market leads to additional downward effects of tightening macro-idential policy. The sign is consistent with our conjecture but it's not significant. In the second column we do not find a significant coefficient for the impact of the growth in household indebtedness in that local housing market but it also has the opposite sign because it's positive. The reason here is that what we are capturing with our measure of household indebtedness which is the amount of mortgage credit the number of mortgage credits per resident it's the extensive margin of household debt. So we do not pick up the intensive margin of household debt and of course that's what you want to capture the amount or the change in risky debt position in your local housing market and that's probably the reason why we don't find a significant coefficient with the right sign. Me and Lara we spent quite a lot of time trying to find a measure of the intensive margin of household debt but that data is simply not available at the micro level. It's simply not available in the credit registry and if the discussant for example has some advice we would be very happy to hear that. That being said what is probably more interesting is to look at the estimates when we combine different interaction variables and here you again notice in the first graph that we find this positive but almost zero coefficient for the hotness of the local housing market and again an insignificant positive coefficient for this indebtedness indicator. For column two what we did there is combining the indicators of financially constrained borrowers being share of credit to young people and the share of highly educated young people within that local municipality. And we find the share of credit to young people to continuously be negative and significant so having an additional negative impact of market potential policy tightening on local house price growth rates but this seems to be contracted by this measure of highly educated young people that are the presence of those people in the local market which is consistent with our conjecture that those are the people who are less constrained. Even controlling for the amount of young people highly educated in that local municipality we still find a negative coefficient of the share of credit going to young people in that municipality. The third column informs you about our different indicators of risky positions in the mortgage market so it picks up the share of first time buyers in mortgage credit in that municipality. The share of credit to young people so young people being between 24 and 35 and the share of overdue credits. For all these indicators we find positive we find negative but significant coefficients even when we combine them. If we then look at column four there we take together all the indicators of financially constrained residents except for share of overdue credit and the share of single households because there's too much correlation with other indicators. We have correlation being consistently above 70% so we remove them in this column. But what you can see here is that if we combine those different terms that we continue to have negative coefficients for each of these indicators but the significance drops or falls away for the share of first time buyers and the share of single parent households. So and then column five takes together all these indicators together with the ones in the first column. So what we infer from this table is that, okay I'll continue. So an alternative approach that we took to try to find out whether there's an impact of activity in housing markets at the local level is to use a quanta regression approach where we can make our beta coefficient depending on the hardness of the local housing market and we do this via quanta regression approach because it allows us to take the distribution of the dependent variable and makes our coefficients depending on that distribution and hard price growth rates means that if you take the left tail of the distribution you're in a cold housing market. If you take the right tail of the distribution you are in a hot housing market. So this allows us to make all of our coefficients depending on being in a hot or cold housing market in a specific year. So I don't have time to go to the details of our results but what we find is that these measures of financial constraints residents continue to impact local house price growth rate in a negative way but along the activity of local housing markets there seems to be different effects. But I guess I don't have the time to discuss this anymore so I will just leave you with my conclusion slide namely geographic heterogeneity seems to matter. There's a heterogeneous impact of national macro financial tools on local house prices and this seems to be different. These effect seems to differ across cold versus hot housing markets. So conclusion that we take from that is that it matters for policy makers to look at those heterogeneous effects because they can have distributional consequences as well. Thank you. Thank you very much, Sabine. Last question. Okay, thank you very much. Thank you to the organizers for giving me the opportunity to discuss this paper. I found this paper very interesting and very timely. Let me very briefly summarize what Lara and Selin did and then give you my suggestions and some observations. So the paper is analyzing the heterogeneous effects of power-based macro-prudential policies targeting the house market. But where is the key heterogeneity we are interested in? Where their high, so it's a municipal level data that the paper is using and they are focusing on whether the municipalities with higher share of risky borrowers later defined indebted households or municipalities with high housing market activity face a heterogeneous impact after macro-prudential tightening in Belgium. So the authors are using the data between 2012 and 2020 using municipal level hedonic prices. They use the macro-prudential policies taken by National Bank of Belgium 2014, 18 and then recent loan to value that service to income ratio and that income ratio adjustments in 2020. The analysis is municipal level. So all these heterogeneities when we talk about the risky borrowers when we talk about indebted households all the heterogeneities going through the share of risky households, for example, defined as share of young households, single-parent households or low-income households or when we talk about the indebtedness it's again share of overdue credit or share of outstanding credit at the municipal level. And the paper also tries to control for again using municipal level data in terms of demand and supply. So the basic motivation is first of all, we all know that the housing markets are segmented. There are local dynamics. There are lots of local factors affecting the housing prices. And then the question is in response to macro-prudential macro-recording policies in general it's made fiscal policy or in particular case of macro-prudential policies are there also local dynamics due to these different characteristics. But it's also related to the recent discussions about whether the macro-recording policies and in particular macro-prudential policy leads to some distribution effects through the focus on the housing prices. So the basic model is, so the papers addressing only the annual changes in the hedonic house price indices where it controls for factors affecting demand and supply but the key thing there is the beta coefficient that they're trying to estimate. And I use slightly little difference notation. It is the coefficient in front of the intensely adjusted macro-prudential policy that Selen described in detail and the local market conditions that I mentioned a moment ago like share of young households, share of low-income households, share of indebted households and the key hypothesis is whether this beta is in a way that suggests that tightening leads to a larger decline in the housing price growth in areas characterized by risky or indebted households plus where the labor housing market activity is already high. And they use for municipality time and regional time-fixed effects. The key findings is yes, they do find heterogeneous impact. In particular, they see their observed higher decline in house price growth in locations with higher share of low-income households, young people, single households, first-time buyers and overdue credits. And later with the Contile Regression Analysis that Selen didn't have time to show, they showed that the effect is more pronounced in the regions where, which was already experiencing high house price growth. Okay, so what I take out of this thing, first of all, I find this analysis is a novel and a timely analysis contributing to the recent discussions about both the effect of macro-prudential policies and the distributional effects of macro-prudential policies. When it comes to the distributional effects of macro-prudential policies intermediated through the housing market, we have only a handful of papers. Two of them, one of them is a recent paper by Daniel Chariah and The Quarters, published in Journal of Finance and Jose Luis Pedro and The Quarters, published in 2020, just published, I think in September, a review of financial studies. The first one for using Irish data, the second one using the UK data. The paper considers, which in all these constraints, carefully what may be the demand factors and supply factors at the municipal level and also regional level time varying heterogeneity. And the results provide encouraging results in the sense that the macro-prudential tightening may tame the housing price fluctuations. All we have so far are mostly the cross-country analysis. For example, paper by Ken Katrin, Ilio Xim or Eugenio Gerriti, Stein-Kleissens, Luke Lavan. All we have mostly is the cross-country analysis. So using the data for Belgium, the paper suggests that the macro-prudential tightening may lead to slowdown in the housing growth. And it may motivate us further to understand whether and how to design and implement further macro-economic policies if there is any inequality issues as a result. But let me share some of my observations and suggestions. If the questions about the effectiveness of the borrower-based macro-prudential policies, okay, I used to be at the central bank, and the first thing I would be asked was, okay, first show us whether this led to a slowdown in the credit growth, mortgage loan origination, or maybe higher cost of borrowing for the individual. So the paper directly focuses on the house price dynamics. But in my opinion, to give a better narrative and to give a complete picture about the transmission mechanism, we need some more in terms of whether this tightening lead to changes in the credit conditions. And ideally, like Acarya and others paper, or Pedro and others paper, this, the paper would benefit from loan-level data. I don't know very much about the Belgium data, but the municipal-level data that we have here, maybe silent, I'm still, I will have some suggestions, but the ideal thing would be, I think, to use the loan-level data and to first address if this tightening lead to changes in the credit conditions. And another big question is, I mean, the macro-prudential policy actions are about also changing the risk composition. Here, we observe that the areas with the high-risk households are facing differential extra slowdown in the housing prices in response to macro-prudential tightening. But we don't know whether it is a change in the price or whether it's some sort of a decomposition where the banks now are preferring relocating mortgage loans to, let me say, less risky borrowers. So we don't have this thing in the paper. And from the policy perspective, it's an important thing, I think, to observe if the macro-prudential policy leads to some sort of changes in the composition of the risks. And even without the loan-level data, I think the paper can still make use of things like, as far as I understand, there is the municipal-level data about the growth of housing transactions, growth of mortgage credits. Still, the paper, before starting with the housing price, maybe it can show how the housing transactions are mortgage credits respond and whether that response is differential with respect to riskings in the municipalities. Some further observations and comments. The empirical specification here, I wrote here one year, but instantaneous impact. The empirical specification assumes, but Lauren Selin has another paper with a cross-country analysis in which, what I observe is the significant effect starts from four to six quarters, but the main effect goes beyond four to quarters. So maybe it may be good to reconcile the leg structure or the outside leg of the policy. Okay, there are some econometric things, maybe I can share during the COVID break, for example, and some of them has been fixed in the slides compared to the earlier version, but one thing also linking to a discussion yesterday. The paper starts with all saturated fixed effects, but I think the nice thing would be to remove, let's say, municipal fixed effects, other things, and to see how these things, for example, the share of high-risk individuals in the region are correlated to the health prices and then introduce the fixed effects one by one. So there are some other comments, I think for the sake of time, I can give these things during the thing. In very short summary, it's a timely paper, it's a novel paper in an area where there's still a knowledge gap, and the paper has potential to contribute to the recent discussions on the effect of some macro-prudential policies and also the dynamics of housing wealth and inequality, but in my opinion, there's a clear potential to benefit from more granular data at the long-level day, which would tell us more about what really the transmission mechanism is at the end of which we observe house price responses. Thank you so much. Thanks very much to both. So you will have the chance to answer the questions. We can take maybe one comment, and Gürgen was the first, so. Thank you very much indeed. It's very timely, very useful. I'm a believer in micro-approaches, so this isn't very timely. One quick comment is whether, in the same way as you looked at the quantile, with the quantile approach at hot versus cold, have you considered or could consider spatial dependence models in which basically, the municipalities actually may have some interaction effects with the neighboring municipalities because people may also live in one work in the other and then may also influence the demand for temporary housing across various municipalities. And a bit more provocatively, also having in mind what Marco yesterday commented in terms of, okay, we go to border-world-based measures for NFCs and the order of magnitude in terms of complications for calibrations. What does this mean? What would this mean in your opinion for calibrating border-world-based measures, for example, at a more local level? So thanks very much. We can allow for one other one. There was the lady over there, yeah. Thank you. Yanar Dhanoa from the Danish Central Bank. I have a different policy question. So in your paper, you combine the capital-based measures with the borrower-based measures. I was wondering if you've considered having them separate, basically, because in the literature otherwise, there seems to be a very different effect from capital-based measures compared to borrower-based. Capital-based measures are generally not considered to be having an impact on prices or these dynamics, whereas borrower-based measures are considered really effective with a large pass-through. And also having these dummy type of variables, it seems that the pass-through from the measures you consider are the same. Also along the same line, it seems that the borrower-based measures were introduced at the end of the period that you're considering. So I was just wondering whether the results could be interpreted as only the effects from the capital-based measures and not so much from the borrower-based measures. And just finally, I was wondering if you've considered to include other policies that might have an effect on the results that you get in the sample. For example, throughout this period, a lot of capital regulation is being faced in. So how do you distinguish between other types of capital regulation and the macro-potential measures? And also, is there any changes in the tax system that would also impact prices? Thank you. Boris Joseline, please be concise. Okay, concerning the question of spillover effects, we've gotten this question quite a lot already before. The short answer is that at this moment, we don't take that into account. But if there is a way to do that, we would definitely want to, because we could think about sizable spillover effects. Just looking at the data, we know that people in Belgium don't tend to move that often and that far away from their parent's house. But still, this is a limited and very restrictive assumption. Concerning your question on what that would entail for a calibrating firm measures at the local level. Yeah, I think it's already quite complicated to take into account all the different segments of the housing markets with young people, first time borrowers, buy to let investors, and so on. That's manageable in terms of policy, but I agree with the discussion of yesterday that in terms of firms, there's even a lot more heterogeneity beyond the location. And that's, we are typically focusing in this paper on the location of housing markets as a driving measure of segmented markets. And then concerning the questions, there are a lot of questions, which are very interesting. I cannot answer all of them, I guess. The thing is that we've indeed combined capital-based measure with borrow-based measures. Yeah, the main reason to do so is because there aren't that many measures being undertaken. Our idea is to get an update of the data. We are waiting for this hedonic house price index. Then we have an additional observation of changes in capital-based measures in Belgium and maybe then we can limit it to the capital-based measure. We can estimate them separately, but at this moment it seems that we don't have a sufficient amount of observations, which is always the issue in empirical work relating to my credential measures. But we are thinking about that. So thank you for the suggestions. Other measures at this point have not been taken into account because we really wanted to focus on those measures implemented to address the resilience to the volatility in housing markets. But we can take that into account because we do have them from our order break. Okay, thank you very much, Celine and Sona. So let's move now to the second paper in this session, again on housing markets. And I would give the floor to Alexander Varadi. She's senior research economist at the Bank of England and has a research interest, mainly in the household behavior and housing market. Area and the paper will be discussed by Christoph Basten from the University of Zurich, where he is assistant professor of banking and is working on macro-potential capital requirements among other topics. Hello, everyone. Thank you very much to the organizers for inviting me to present my work today. So the work I'm going to present today is joined with my colleague, Bruno Albuquerque, who's an economist at IMF, and the usual disclaimers apply in that the views in this paper represent only our own and not the views of the Bank of England or the IMF. So I wanted to just start by setting out our kind of research questions for this paper. So the aim of this paper is to try to examine the ability of mortgage payment holidays that were introduced in the UK during the COVID-19 pandemic to help support household consumption. And we have three key questions that we're interested in answering in this paper. So first, we want to understand if mortgage payment holidays have been accessed by the households who need them the most. Then we want to understand how have mortgage payment holidays actually supported the consumption dynamics of the mortgage orders. And then third, we're interested in the distributional impact. So we want to understand if mortgage payment holidays have been more potent for those households facing threats to their financial resilience. And the motivation for this paper is twofold. So first, we want to contribute to the literature on the effective government policies on households. So existing research in this space, in the space of kind of policies that offer temporary liquidity constrained to households has been mainly focused on policies such as direct cash stimulus or tax rebates in the US. The impact of mortgage payment holidays per se on kind of consumption in recessions has not been studied to our knowledge. The policies have been implemented in some other form in the US as well. But in the US, existing work during COVID has been focused on the impact of these policies on mortgage defaults rather than on spending patterns. The paper that's most similar to ours is by Villala in 2019 who does study the impact of mortgage payment holidays in Finland in 2015. But implementation of these policies in Finland were very different to what we observed in the UK. So in Finland, borrowers were still liable to make interest repayments. While in the UK, we had a complete stop to both capital and interest repayments during the recession. In Finland, the economy was still in a boom when the policy was introduced. Whereas in the UK, we considered the impact of a policy during a large aggregate negative shock. And in Finland, the policy was offered by one lender only. Whereas in the UK, it was completely nationwide. Our second motivation for this paper is obviously for policy. During the last financial crisis in 2007 to 2008, arrears of mortgage orders increased drastically. And spending particularly of highly indebted households decreased substantially, which led to kind of amplifying the negative shocks of the crisis. However, during the COVID pandemic, with mortgage moratorium in place, and of course with other government policies to support the real economy, we have seen arrears remaining low. So then there is an important question of whether mortgage payment holidays can be used during an aggregate negative kind of shock to deal with an aggregate negative shock. And whether policy makers should consider this policy going forward. So I wanted to spend some time just on kind of explaining how this policy was implemented in the UK during the COVID-19 pandemic. So as I mentioned, it was a nationwide government policy that was introduced in March, 2020. It offered an initial break for three months, which was then extended to six months in June, 2020. And it offered a complete stop in both capital and interest repayments. It was free and easy to apply for the policy and the government kind of told households that there would be no consequences to their credit scores. Approximately 20% of all mortgage orders in the UK took a mortgage payment holiday and the majority of these payment holidays were extended at the start of the pandemic. The chart here shows both the flow and the stock of mortgage payment holidays. And as you can see by the red line, the majority of mortgage payment holidays have been extended between March and May, 2020. The stock had slowly grown up to a peak of around 17%, which was reached around May, 2020. And then it kind of slightly started to decrease. So I wanted to kind of spend some time on the data that we use in this paper. So the data that we use comes from a mobile app called Money Dashboard. The app provides a free real-time account aggregator for users that collects all users' daily transactions from their current accounts, credit accounts, and savings accounts within one single platform. It includes information on the user's age, gender and postcode, and from the data that we have, we're also able to estimate the presence of kids. So in terms of kind of the sources of funds for the users, we are able to observe income. And we're able to observe whether this comes from salary, whether it is rental income, financial income, or comes from benefits, including government benefits. And we're also able to observe unsecured debt finance. So this is kind of personal loan finance, for instance, coming into users' accounts. On how users actually use their funds, we can observe spending at the very, very granular level that we then kind of aggregate into different buckets. We can observe cash withdrawals, payments into investment accounts, rental payments, mortgage repayments, as well as unsecured debt repayments. After cleaning, we have around 13,000 distinct users which we can observe on a monthly basis from January 2019 until November 2020. So we do a lot of testing in the paper to try to show that this transaction level data that we get from Money Dashboard is indeed representative of mortgage wars. So on this slide, I just plotted two of our kind of key charts showing how representative the data is. So starting with the chart on the right, that plots the percentage of users in Money Dashboard who are mortgage wars, and compares that with two other sources from the government. One that comes straight from the government family resource serving in the UK, and one that comes from the Office of National Statistics. And you can see that Money Dashboard does a really good job at identifying those households who are indeed mortgage wars, so for whom we can observe a monthly mortgage payment. Where Money Dashboard is not that good is identifying which households are renters and which households are outright owners. And that's kind of understandable because transaction level data sets like Money Dashboard use an algorithm to try to track transactions and to try to kind of identify where those are rental payments or mortgage payments. And it's quite hard to identify rental payments because those are essentially just transactions between two different households. And it's quite hard to know whether there's just payments for random spending or whether that's the trending. However, for the purpose of our paper, we don't necessarily put a lot of weight on that because we do focus in this paper on mortgage wars and we bucket everyone else renters and our right owners into one category. The chart on the left plus the distribution of income for mortgage wars in Money Dashboard alongside data from the Office of National Statistics in the UK. And you can see that the distribution is actually not that bad. So with the exception of one of the tails, Money Dashboard does a pretty good job at identifying kind of income across the data. So I wanted to kind of spend a bit of time on how we identify mortgage payment holidays in Money Dashboard. That's because we don't have a dummy variable that switches from zero to one if a mortgage payment is missing from the data and that's because of payment holidays or something else. So in order to identify mortgage payment holidays in the data, we have an algorithm in three steps. So in step one, we identify consistent mortgage payments. That is mortgage payments that we can observe on a monthly basis across multiple months. And that's because sometimes households make early repayments, for instance, there are very sporadic and kind of different amounts. And we wanna make sure that we don't tag those as kind of mortgage payment holidays. Then in step two, we identify a missing payment as a mortgage payment holiday if it's missing from March 2020 when the policy was introduced, but it was observed previously or if it had been missing for a few months prior to March 2020, but it then resumed later in 2020. And the idea is here to try to capture households on our rears as well. So at the very start of the pandemic, the government allowed households who were in rears prior to the pandemic to also jump on mortgage payment holidays because it was more beneficial for their credit scores. We wanted to capture those mortgages as well. And then third, we have to kind of see the payments resumed by November 2020 when our data finishes. The chart here shows the stock of mortgage payment holidays that we identify with this algorithm from March 2020 until October 2020. And it plots the stock of mortgage payment holidays in money dashboard against data from UK Finance. So the UK Finance is a body in the UK that provided during the pandemic aggregated data on mortgage payment holidays from a sample of lenders. And as you can see, we do quite well with our algorithm in identifying mortgage payment holidays that it's kind of very representative for what the lenders themselves were saying at the time. So moving on to kind of to analysis, one thing we wanted to look at in this paper is to kind of try to understand who were the mortgageers who access mortgage payment holidays. And in order to do that, we run a probit regression on the sample of mortgage yours where we're trying to understand the correlations between the probability of getting a mortgage payment holidays and household characteristics. And in terms of household characteristics, we consider households financial position prior to the pandemic. So we look at households that service ratios prior to the pandemic, their savings rates prior to the pandemic, whether they had financial income going into the pandemic or not, whether they had unsecured credit entering the pandemic and the number of mortgages that they made on a monthly basis prior to the pandemic. We also look at the financial position during the pandemic. So we look at whether their income has changed at the very start of the pandemic relative to the months before. And we look at whether households were made unemployed at the very start of the pandemic compared to their status prior. We also look at household characteristics. So we look at whether households have kids and we also put age to the regression. So what we find here is that there is a higher probability of policy take up if households were entering the crisis with existing financial vulnerabilities and in particular with high debt service ratios or low savings rates. And this is somewhat expected. We also find that for households experiencing negative income shocks at the start of the pandemic, they were also much more likely and they had a higher probability of policy take up and they were much more likely to have a mortgage payment holiday for a period longer than three months. What is interesting with this analysis is that mortgage payment holidays didn't go just to the people who were most kind of financially strained. It also went to people with a better kind of financial position. So we find that households who had a positive financial income entering the pandemic as well as households were more likely to be property investors who had second and third homes, they were also much more probable to have a mortgage payment holidays compared to households with other characteristics. So moving on, we kind of, we want to understand, okay, so for households who had mortgage payment holidays, what would the policy actually do to their spending patterns? And in order to answer this question, we apply a difference in difference approach as shown in the regression in the slides. So we regress the year-on-year change in the logarithm of real non-housing consumption on a number of controls that I'll explain in a minute and on dummy for whether households had a mortgage payment holiday or not, also interacting that dummy with some household characteristics that again I'll go through in a minute. So the coefficient of interest here is beta zero, which captures the difference in average consumption due to mortgage payment holidays between the treated and the control group. And the kind of, the main issue with this regression is that it could be very prone to both self-selection into policy because the policy was offered to all the mortgage owners in the UK, but there's also many sources of non-randomness that we try to control for. So first, we try to address these issues by choosing the control group to be the non-eligible for the policy. So renters and outright owners will be our control group. The reason why we don't choose mortgage owners who are never on mortgage payment holidays to be our control group is particularly because of self-selection. So the policy was offered to everyone, but only 20% actually chose to take it up, which means that there could be many unobserved factors amongst mortgage owners that may determine policy, the probability of policy take up, which we are unable to control for. And one example here is financial literacy. So although the policy was very highly advertised across many different media outlets, it's still possible, according to the Financial Conduct Authority in the UK, that a big chunk of mortgage owners didn't know that it existed, so they never applied for it. This is not something that we can control for, but this would be something that would bias our estimators if we were to kind of choose mortgage owners as our control group. We then include a wide range of fixed effects and other controls into the regression. So we control for time varying income, as well as for changes in income for households at the very start of the pandemic relative to the months before. This is what's in the slides labeled as income shocks, but we also include fixed effects for households, region, country, postcode and time. And then finally, we do a wide range of robustness checks just to make sure that our estimators say we remain consistent. So the first thing that we do is we test two alternative identification strategies for our coefficients of interest. I won't have time to go through them in a lot of detail in this presentation, but I just wanted to mention them here. So the first one is synthetic control methods where the control group is computed by weighing up the control units that mostly resemble the treated units in terms of the consumption patterns prior to the pandemic. The second alternative approach is propensity score matching where we match controls and treated units based on observables, many observables. And compare outcomes for treated and those controls that match the treated units the best. We also try a robustness check where we use mortgage orders for never on mortgage payment holidays as controls just to see if our results would be wildly different if we were to choose that control group despite our self-selection issues. And we also allow for anticipation and delayed effects. We allow households to anticipate the policy and start making changes to their consumption prior to actually accessing the policy as well as to kind of delay the effects of the policy. None of these robustness checks change or estimated coefficients, they remain very, very robust to kind of all of these approaches. So kind of moving on to results. So starting with kind of column one, the first thing to note here is that mortgage payment holidays are not statistically significant for the average mortgage order. So the average mortgage order on mortgage payment holidays did not increase their spend or did not benefit from higher consumption just because they kind of use the policy. However, when we look into the sample of mortgage orders for liquidity constraint, the results change. And we do capture liquidity constraints in various ways. So in column two, we include a dummy variable that equals to one if households are in the lower quintile for their savings rates. In column three, we include the savings rates linearly in the regression. And in column four, we include the dummy variable for whether households are in the kind of bottom quintile for the income distribution. And for all of these, we get kind of positive and significant impact of the policy. In column five, we also look at whether households who have withstand a negative income shock at the start of the pandemic consume more relative to the control group who had suffered from similar income shocks. And once again, we get a positive impact of the policy. Now, the specification that we most like is the one presented in column seven that kind of says that the following things. First, amongst households who are more likely to be liquidity constraints, so those in the bottom quintile of the savings rate, those mortgage orders were on mortgage payment holidays benefited from a consumption growth that was around 22% higher than the control group with similar liquidity constraints. And the second result is that amongst households who suffered from a 10% negative income shock at the start of the pandemic, mortgage orders who were on mortgage payment holidays benefited from consumption growth that was 0.3% higher than control group with similar kind of income shocks at the start of the pandemic. I just wanted to quickly highlight that our results stay the same with one of our alternative approaches, identification approaches. And I've shown here the propensity score matching. So as you probably know, the propensity score matching is done into stages. In stage one, we estimate the probability of policy take up across treated and non-eligible given a wide range of observables. And then in stage two, we match treated and controls based on propensity scores using the nearest neighbor estimator. The first row shows the average impact across the sample. And once again, you see that there's no statistical significance effect of the policies. But when we look at the sample of liquidity constrained households only, then propensity score results plotted in column one and two are very, very similar to results that we get out using our different base line showed in column three. And before I finish, I just wanted to kind of show the answer to the question of kind of mortgage payment holidays have not been targeted in the UK. They've been offered to all households with different financial resilience. So what did households who were not liquidity constraints who didn't have important financial resilience risks, what did they do with mortgage payment holidays? And to answer this question, we do the, we kind of rerun the regression, the different regression that I've just presented. But instead of having consumption growth as a dependent variable, we have the savings rate. And this is plotted in column two with column one showing the different results when consumption growth is used as a dependent variable. And as you can see for the average mortgage or consumption hasn't changed, but their savings rates have increased considerably. And amongst the liquidity constrained households, it was only consumption that was statistically kind of affected by the policy with their savings rates not increasing or not changing in a statistical way. So to conclude in this paper, we use rich transaction level data to investigate the effective government policies on mortgage orders. We use many robustness checks and controls to eliminate many sources of bias, including from self-selection from the broader effects of the pandemic and from different government policies that were used at the time to support households. We find that mortgage payment holidays have been used by households with varying degrees of financial strength included by property investors. But we find that payment holidays did not lead the average mortgage or to change their consumption. Instead, the policy supported consumption of households with low savings rates which are more likely to be liquidity constrained. And this is robust to many model specifications. Thank you very much. Okay, thanks a lot to the organizers for giving me the opportunity to read and discuss this exciting paper. Okay, I will provide a very brief summary, just two slides and then I have up to six comments depending on how much time I have. I hope some of them are useful. Okay, brief summary. So the office evaluate UK mortgage moratoria where during the pandemic, households could pause repaying the interest and repaying the principle of the mortgage. And they look at three, in my reading, they look at sort of three outcomes first. They look at who takes up, who uses this policy. Second, for those who do use it, how does it affect their consumption growth? And if they don't consume it, then obviously the alternative is savings. So how does it affect their savings rate? They use transaction level data from the UK's biggest aggregator app called Money Dashboard and I see two key findings. One is the payment holidays do allow the liquidity constraint households to smooth their consumption. The other thing is the payment holidays are also used by households who are not liquidity constrained. And then we have to discuss, is that a loss somehow of efficiency or does it not matter? The basic methodology is diff in diff. So they have an interaction between a dummy for using the payment holiday and between different measures of liquidity constraints. Okay, that's my summary. Now, what comments do I have? The first one is when it comes to measuring effectiveness, they use different measures of income shocks and they use different measures of debt. I would have used the debt service ratio, somehow the ratio between the two because if you low income, if you have low savings, arguably that doesn't matter much if you don't have much debt or maybe you have super high debt but it doesn't matter much because you have an income to back it, right? So I would use the debt service ratio. Also, I would not use it just pre-pandemic and pre-payment holidays. I was thinking of using how the DSR changes over the pandemic if you have an income shock. So you have the debt before when your income is shocked and how does that affect your take-up and the treatment of the payment holiday? Obviously, one problem potentially is that it's endogenous. It could be that if you have a big mortgage, maybe you work harder to not have your income drop. Maybe if you can afford it, you say, well, I work less. So I was thinking, is there a way to instrument the income shock? And some things I thought of is would you look at the sector of employment? I understand you don't have the data. I guess in the app, you see income coming in. So I don't know if you can look at the text and see where the employer is. But what I understand you do have is the postcode. So the region could maybe be used as an instrument. For once, different regions were differently affected by the pandemic. They had different sectors, right? So a region where there's a lot of hotels and restaurants might be affected more. Maybe that makes you lose your job. If you work in IT or research, maybe it doesn't. The other thing is the region might affect who else takes up the payment holiday. So there is this paper by Guiso Sapienza and Zingales showing that if you have more peers in the US who default on their mortgages, that also makes you more likely to default. So here it could be peer affected. It could be knowing about the payment holiday. So you write, there's only 20% take up. Maybe not everyone knows about it. But if you are in a region where many others use that opportunity, maybe you're more likely to take it up. So that's my first comment. Relatedly, so you say you can't use the non-mortgages as control group and you discuss who takes up the policy. And you do find that it's more likely to be used by those with higher ex-ante debt service ratios. And I think that's very good. Yeah, and as I just said, I think I would also look at the regions where people live and I would not only look at a realized unemployment shocks but also at unemployment shock risk. So I would imagine if you work in hotel or restaurants, you're more likely to lose your job. So even if you don't observe people losing their job, I would still think they might, for example, take up the payment holiday and save more because they are anticipating that they might lose their job. Okay, third point of interest is the control group. So you write you don't use existing mortgages as control group. I understand from some of the texts that you used that earlier and you also say, said in the presentation, you use that as a robustness check. So the main control group you use is renters and outright owners. I was a bit surprised that you say that only 30% in your sample borrow because the home ownership, as I understand, is about 60% in the UK. I Googled it, I found some slightly higher numbers, 38%, but what I was wondering is if there's so many homeowners in the UK and only half of them are mortgages, I take it that people repay faster. So I'm more used to the Swiss or Dutch setup where people just don't repay for tax reasons. But if people repay fast, that's maybe an opportunity because maybe you can look at mortgages who have repaid a year or two ago. It's still the same kind of household who buys a home, who takes up debt, but maybe if they have repaid last year that helps. So there's cross-sectional differences, obviously for different, if you're interested in the trends. The trends are parallel in your graphs before the treatment. They also look pretty parallel to me after the treatment. So I was wondering, are you plotting the right outcome? And so as I see it, you plot consumption levels which are parallel ex ante, but in the regressions you use consumption growth. So I would also plot pre trends for consumption growth and see how parallel these are. Maybe they are not entirely, that's fine, but then you just wanna discuss it. Okay. Yeah, there's one issue of the extent to which the app captures all transactions of a household or whether there's accounts that are not linked. I'll skip that. But the second point is how selective is the app use? You say it is representative of the population. I would still think maybe it's more the young or the more educated who use it. And maybe you wanna discuss if there are any biases. I don't think that makes the paper wrong, but maybe that means that your estimates are, if anything, upper or lower bound. So I think in that way it could be useful to discuss that. Outcomes, you say you drop cash because you don't know what people do with it and you drop un-tech debit transactions. I would have included both. I would have thought it's unlikely that people withdraw a lot of cash and put it under that pillow. I think most cash will be used for consumption. I would just have included cash and un-tech debit transactions as consumptions. Okay, the final thing we're discussing is the interpretation. So you said the FCA, the regulator recommended that banks give payment holidays, but I was wondering what are the incentives for the banks to do so. So I imagine maybe they think that reduces the default risk. If those who can't afford to pay, get a holiday, maybe they get goodwill because the mortgages say, well, I like my bank because it helps me. So I was wondering, is there any cost? That's the third point. Is there any cost of take-up of non-liquidity constraints also? So when there's an outright subsidy or if they don't have to ever repay, obviously there's a huge cost. If it's just paused, I suppose the banks lose some interest. I wasn't clear from that. If I repay six months later, do I have to pay more interest? I guess in a low rate period, that's not a lot of cost. I guess if banks only have liquidity risk because the payment comes later, probably there's other measures to capture that, but it would be good to discuss whether there's any cost to the take-up by non-liquidity constraints or whether that's totally fine. Last point I wanna talk about is point two. I was really wondering what are the effects of the payment holidays after the pandemic? So what you do discuss is how does consumption change just the month after the payment holiday? But if you have the data, I would find it very interesting to see what do they do a year or two after the pandemic because after the payment holiday, they might still be worried that the pandemic might get worse. They might still want to do more saving, but I would find it interesting what do they do like a year afterwards. That's the main points. And yeah, thanks for the exciting paper and I'm looking forward to seeing a future version of it too. Thank you. Thank you very much. We can take a number of questions. I have one over here, there. Hello, thank you for the presentation. I was wondering if you have a sense of how many of the people who could benefit from this policy actually took it up? So kind of to reverse the regressions a little bit because I think that's really important from a policy perspective. Thank you very much for the very interesting paper that uses novel kind of data. I wonder if you took that extra step. You said that liquidity constrained borrowers are mostly smooth their consumption after this payment holiday. I was wondering whether you did or would aggregate the effects of both liquidity constrained and unconstrained and estimate the total impact of consumption and talking about the wider costs of this payment holidays in terms of what the discussant discussed, I was wondering, you know, is this payment holiday as an exposed macro prudential measure upon a sale? Basically, it's the way I think about it is supporting credit, avoiding a credit crunch and also supporting consumption. But during the Corona crisis, the banks were basically not liquidity constrained. So you're basically supporting liquidity constrained households but banks do have liquidity. What about in other cases when banks are themselves liquidity constrained? So what would that kind of larger macroeconomic effects entail? Would there be a credit crunch or not? So that's just a general idea. Thank you. Thanks very much. Okay, then over to you, Alexandra. Thank you very much for your comments as well as for the questions. I'll start with the questions if that's okay. So the kind of the first question, I think it's very interesting is how many people actually took the policy amongst all the liquidity constrained. So in money dashboard, about a half, but we don't know in reality. And I think that's kind of, that's quite important. We do show that the data is representative but it's just kind of a sample of the data that we see of 13,000 households. So it's quite hard to aggregate that at the kind of national level to see how many actually benefited. But in money dashboard, we see that about a half of the people who were liquidity constrained actually to the mortgage payment holidays and so benefit to consumption. The reason why others may not have taken the mortgage payment holidays is, as I mentioned in my slides, very likely financial literacy. So there is a paper from the Financial Conduct Authority that looked at different characteristics of household or mortgage orders during the pandemic and has found that around 20 to 30% of those who would have benefited from the policy didn't actually use the policy because they just weren't aware that it was there and it was based on essentially survey data. So we think that although it was very much advertised across the UK, it just didn't reach everyone. And then in terms of the second question about the total impact on consumption, we take that. I think we think it's very important. It's just the data that we had didn't allow us at the time to study this fact, but it's a very important one and it's the one that most matters for macro potential policy, particularly in the UK, given that we say that one purpose of macro potential housing tools in the UK is to avoid this aggregate demand externalities, the fact that highly indebted households tend to pull back from consumption a lot more in bad times amplifying effects. So we don't have a way of answering that at the moment. We are trying to get updated data from money dashboard in order to try to assess the longer term impact of consumption on households. Now, one thing that we do in the paper is we look at what happens at the expiration date. So for households on both three months and six months for which we can observe the date at the expiration date, what we do is we say, okay, in the two or three months following the expiration date, do they, their consumption, is that still positive or negative? And what we find is that for households on three months mortgage payment holidays, there are no statistical significant effects of consumption. So essentially they consume more while the policy is active, but then when the expiration date, that there's no statistical significant of spending anymore. However, for households on six months mortgage payment holidays, we do find that they tend to over consume during the, well, the policy is active and they tend to under consume at the end. So we see a statistical negative effect in terms of consumption in the two to three months following the expiration date. And we think that's mainly because households who are on longer mortgage payment holidays, they tend to be correlated with income shocks as well. So they tend to be the ones that have suffered a large negative income shock at the very start of the pandemic. So they are even more kind of financially in a difficult financial place in addition to kind of having no savings raised to deal with the pandemic. So they need a little bit longer to have the mortgage payment holiday for, but they also might be the ones who are most constrained and they're not able to consume optimally. So they essentially increase consumption while the policy is there, but then they have to pull back from consumption at the end of the policy. And I don't know if I have time to answer the, yes. So just to answer a couple of comments from Christoph, so one is on the sample selection. So you mentioned that it's on one hand, it's a bit unclear what household is in the context of our paper. And what we do is we use the concept of nuclear family. So we assume that some users may kind of tag their partner's accounts into their own app and that is their decision. And if they do that, then we kind of take the household as a given for households who don't do that and kind of we see only one user's accounts, then we just take it as given. It's not the best approach, but we decided to kind of take all the data that we have rather than kind of cut and say, we're only looking at single households instead. In terms of the data being slightly biased, that's a fair point. So the data is when we look at the whole households, money dashboard is biased towards young and educated people, mostly residing in London and the South. However, when we look at the distribution across mortgageures, this bias is not necessarily there because mortgageures in the UK tend to be younger, educated and residing in London. So that is why we kind of focus our analysis on mortgageures rather than kind of looking at households as a whole. And in terms of kind of the interpretation, you asked about the incentives of banks of doing so. I mean, we have some bankers in the audience so maybe they can chip in as well. But as far as we've seen as regulators, it's mainly because of default. So defaults are very costly in the UK and elsewhere and being able to just help households whether through a health shock that was in kind of no way their own fault sounded cheaper than kind of letting them default. And indeed during the financial crisis, mortgage moratorium weren't in place in the UK. Many households did indeed default. And there was a notable impact on consumption as well. So kind of trying to avoid that loop seemed way less costly by a mortgage moratorium than defaults. And I think I've answered your question on the longer term impact on the pandemic as well, which I think it's a very well taken point. Thank you. Thanks very much, Alexandra and Christoph. So it's time now for the lunch break. We will, the lunch will be served outside in the foyer right outside this door. And we will resume it to o'clock but let me finish with a final applause for the four presenters and discussions today.