Household optimism and overborrowing (cid:3)

We use Finnish household-level data from 1994 to 2013 to measure how often and what kind of forecast errors households make and how the errors are linked to the households(cid:146)borrowing behaviour and overindebt-edness. We (cid:133)nd that those households that make the largest optimistic forecast errors have greater debt-to-income ratios. They also are more likely to report that they su⁄er from excessive debt loads and have problems in coping with their bills. There are no such systematic e⁄ects for the households that make pessimistic forecast errors.


Introduction
Conventional view in economics is that households use unsecured (Sullivan 2008)   and secured (Hurst and Sta¤ord 2004) debt to optimally smooth their consumption over time (see, e.g., Alessie and Lusardi 1997). 1 On the other hand, there is evidence that households may borrow imprudently and accumulate too much debt.These behaviours appear to be related to behavioral biases and poor debt literacy (see, e.g., Stango and Zinman 2009, Campbell, Jackson, Madrian and Tufano 2011, and Lusardi and Tufano 2015).This paper shows that households' indebtedness and unsustainable levels of debt are strongly linked to optimistic forecast errors that the households make.
Our data are a rotating, nationally representative panel that covers a large number of Finnish households over a nineteen-year-period from 1994 to 2013.
These data are particularly suitable for an analysis of the relation between households'forecast errors and their indebtedness.First, the data allow us to measure at the level of a household how its …nancial expectations are related to the subsequent realizations.We use the di¤erence between the subjective expectations and the realizations as a direct indicator of the size and nature of households'forecast errors.Second, our forecast error data are comparable to those used elsewhere in the literature.The survey questions on which our measures rely are nearly identical to those used e.g. by Souleles (2004), underlying the Index of Consumer Sentiment in the US.Third, unlike the other data sets available to date, the Finnish data allow us to directly link households'forecast errors to the same households'borrowing behaviour and perceived overindebtedness. 2 We say that an optimistic forecast error is observed when the expected outcome is better than the realized and that a pessimistic forecast error occurs when the opposite happens.We …nd that on average, households have neither made optimistic nor pessimistic forecast errors over the sample period.The households making the largest optimistic forecast errors have higher debt-toincome ratios.This result is robust to controlling ‡exibly for age, cohort and time e¤ects, demographics, disposable income as well as for the stock of debt cumulated in the past.We show, moreover, that the households that make the largest optimistic forecast errors also are more likely than others to report that they are overindebted and have problems in coping with their debts and bills.
There is a clear asymmetry in how households'indebtedness and unsustainable levels of debt are linked to the forecast errors.The households making pessimistic forecast errors have lower debt-to-income ratios than those making optimistic forecast errors.They do not su¤er from overindebtedness like those who make optimistic forecast errors do.
We also separate households that make "prudent" optimistic forecast errors from those that make "non-prudent" forecast errors.A prudent optimistic forecast error refers to those cases in which a household experiences a negative surprise relative to what it expected but, despite of that happening, it reports that its …nancial situation does not actually get worse (in absolute terms).A household is said to make a non-prudent optimistic forecast error if it both experiences a negative surprise relative to what it expected and if its …nancial situation actually gets worse.To sharpen our baseline results, we show that households that make non-prudent optimistic forecast errors have greater debt-to-income ratios than those who do not make forecast errors or who make pessimistic forecast errors.Those that make non-prudent optimistic forecast errors are also more likely to report that they are overindebted and experience empirical analysis.
problems with their debt than others.We do not …nd similar results for those who make prudent optimistic forecast errors.
Our …ndings contribute …rst and foremost to the scant literature on the economic consequences of households'expectational errors.In an important contribution, Souleles (2004) used the Michigan Survey of Consumer Attitudes and Behaviour together with the Consumer Expenditure Survey to document that in the US, households'forecast errors are correlated with their demographics (and thus ex post biased) and that optimistic forecast errors are negatively related to consumption.Puri and Robinson (2007) used the Survey of Consumer Finance and a di¤erence between self-reported life expectancy and statistical life tables to construct a measure of consumer optimism.Using this measure, they documented that besides heterogeneity in the degree of optimism, there is a negative relation between holding optimistic views and many economic choices, such as saving, balance payment habits of credit card holders, long-term planning and non-smoking. 3r analysis di¤ers from these previous papers in two major ways: First, neither Souleles (2004) nor Puri and Robinson (2007) link households'forecast errors to their total indebtedness or to their problems in coping with payments.
Second, we consider di¤erent degrees and types of forecast errors (e.g., prudent vs. non-prudent).In addition, we show that the borrowing behaviour of those making pessimistic errors di¤ers from the borrowing behaviour of those making optimistic forecast errors.
Our results also contribute to the literature on the …nancial fragility of households.Nearly half of the US households report that they are probably not able to come up with $2000 in a month should a need arise (see Lusardi, Schneider and Tufano 2011).This suggests that already small unexpected …nancial shocks can cause problems to them.In our data, nearly 5% of the households make "large", clearly optimistic forecast errors.The mean drop in annual disposable income among these households is lower than the threshold of 1500 euros used by Lusardi, Schneider and Tufano (2011) to measure which European households are likely to be …nancially fragile.Besides drawing on savings and relying on intra-family transfers, borrowing is one of the most commonly used means to cope with …nancial shocks, both in the US and other industrialized countries (Lusardi, Schneider and Tufano 2011).Our …ndings show that optimistic forecast errors that on average are not extremely large in monetary terms are strongly associated with problematic borrowing behaviour and problems in coping with payments.This suggests a link between the …nancial fragility of households and the type of forecast errors that they make.
Finally, our analysis is related to analyses of debt illiteracy and behavioral biases in the market for debt.Lusardi and Tufano (2015) show that those who appear to have a limited understanding of debt contracts and interest calculations pay more for their borrowing and are more like to be excessively indebted.Stango and Zinman (2009) focus on the systematic tendency of consumers to underestimate loan interest rates (due to exponential growth bias).Gathergood (2012) and Disney and Gathergood (2013) provide complementary evidence on how low …nancial literacy is related to borrowing using UK data.
The rest of this paper is as follows: Section 2 describes the data.Section 3 reports how households' forecast errors are related to borrowing behaviour.Section 4 then shows how the errors are related to measures of overindebtedness and problems in coping with bills.Section 5 o¤ers concluding remarks.

Data
Our main source of data is Statistics Finland's Income Distribution Statistics (IDS).It is a nationally representative data set, covering private Finnish households and their members.The data set is annual and the sampling is based on a rotating panel: Each year, about half are new households; the rest have been included in the IDS once before.This rotation means that each household shows up in the data for two consecutive years. 4e IDS consists of two components: The …rst is a register-based component for which the data are collected from administrative registers, such as census data, tax registers, and social and pension registers.This register-based component of the data contains, for example, detailed demographic information about the households, as well as data on the sources of their income and borrowing be- As we explain in greater detail below, our empirical analysis relies on three key features of the IDS data: First, the survey component of the IDS data allows us to measure expectations.Second, the short panel aspect enables an empirical analysis of the subsequent realizations and, thereby, calculation of forecast errors.Third, the structure of the IDS data allows us to link each household's forecast error to the same household's borrowing behaviour.
The sample available to us covers years from 1994 to 2013.The initial annual IDS sample consists of about ten thousand households per year, but due to the timing of measurement of the expectations variables (see below), there are on average nearly …ve thousand households per year in our data.After allowing for some missing data and non-response to certain key questions and after dropping those households where the respondent was younger than sixteen, our baseline sample consists of 91 827 household-year observations.The exact size of the estimating samples varies a bit, depending on which variables are included.
The sample we use in the empirical analysis is also trimmed for outliers and cleaned for certain data errors. 5 have matched to the IDS data a number of macroeconomic variables, such as the growth rate of gross domestic product (GDP), in ‡ation, unemployment, stock and regional housing prices, and interest rates.The descriptive statistics of the IDS variables and macrovariables used in our empirical analysis are presented in the Appendix (Table A1).
We use the sampling weights provided by Statistics Finland when calculating the descriptive statistics and when implementing our estimations. 6The results are therefore representative of the population.
3 Measuring expectations and forecast errors

Expectations and realizations
To be able to quantify the nature and size of households'forecast errors, we need a measure for household i's …nancial expectation for year t and the associated realization.Our measures of these key quantities are based on the following two debt measures we use in our empirical analysis -at the top 99% percentile of the distribution of the measure.We also dropped those person-year observations for which the change in total debt was at the top 99 percentile or at the lowest 1 percentile, when the debt was measured in euros.There also were some errors in the data, as for some person-years the debt or income measures were negative.They were dropped from the estimation sample(s) as well. 6The sampling scheme overweights entrepreneurs and high-income households.
questions, asked in the survey component of the IDS: Expectation, E itjt 1 , is derived from the …rst of the two surveys in which household i participates.It is based on question "How do you think that the …nancial situation of your household develops during the next 12 months (or during year t)? ".The following response categories are allowed: "1 = is clearly better ", "2 = is somewhat better ", "3 = stays about the same", "4 = is somewhat worse", and "5 = is clearly worse".The question refers to year t, but was asked in the survey that primarily concerns year t 1: Actual outcome, A it , is derived from the second of the two surveys (reinterview) in which i participates.It is based on question "How do you think that the …nancial situation of your household developed in year t? ".It allows the same response categories as the expectation question.The question refers to year t and was asked in the survey that primarily deals with year t. 7 We stress three aspects of these questions: First, the question on which E itjt 1 is based asks about the future development of household i's economic and …nancial position.Its wording and allowed categories match exactly with those of A it , which measures the corresponding realization one year later.These questions are very similar to those used by Souleles (2004) to examine the development of the …nancial condition of US households and their expectations about it (i.e., his variables QF P r and QF P e ).Like Souleles, we use the match between the two questions to analyse in-sample forecast errors at the level of households.Second, the expectation and realization questions line up timewise quite nicely.The survey is typically conducted in the early part of each year.
The forvard-looking expectation question therefore refers to year t and is asked at the beginning of t (in Spring), and the realization question, asked as a part of the re-interview at the beginning of t + 1, looks back at t. Third, it is important to note that while households show up in the data for two consecutive periods (until 2009), we can match E itjt 1 with A it only for the latter period. 8ble 1 displays a cross-tabulation of E itjt 1 and A it .The entries are the number of observations falling in each cell and the associated cell probabilities.
The table shows that a bit more than half of the observations in the data (57%) can be found from the diagonal cells.Moreover, about 34% of the observations are in the cells that are adjacent to the diagonal.These patterns indicate that expectations are strongly, but not perfectly, correlated with the subsequent outcomes.
[Insert Table 1 about here] We can also infer from Table 1 that A it vary around E itjt 1 symmetrically.
This indicates that on average, households have neither made optimistic nor pessimistic forecast errors over the sample period.
The cell probabilities of the table also indicate that the fraction of households who appear to make larger forecast errors in either direction is moderate but not negligible.For example, out of those who expect that their …nancial situation does not worsen (i.e., those for whom E itjt 1 = 1; 2 or 3), around 14% …nd that it actually gets worse (i.e., A it = 4 or 5).

Measuring forecast errors
The matched pair of questions, E itjt 1 and A it , allow us to calculate a forecast error for each household.We consider two complementary measures, which combine insights both from Souleles (2004) and Puri and Robinson (2007): Like Souleles, we use the di¤erence between E itjt 1 and A it as the basis of our forecast errors (see, e.g., his "F P measure).However, instead of just focusing on the di¤erence, we follow Puri and Robinson and explicitly consider the nature of the forecast errors, such as whether the expectations and realizations imply mild or extreme optimism or non-prudent expectation formation.
Our …rst measure, denoted F E1 it ; allows for 5 categories.They are clearly pessimistic forecast error (F , and clearly optimistic forecast error (F E1 it = 5 if The logic of this measure is that it characterizes whether the expectation of household i about the development of its …nancial situation matches with its ex post view of the eventual realization.The optimistic (pessimistic) errors refer to the cases in which a household experiences a negative (positive) surprise relative to what it expected.For example, a household is said to make a clearly optimistic forecast error (F E1 it = 5) if it initially thought that its …nancial situation would clearly improve, but if it in the end stayed about the same or worsened.
It is also worth pointing out two features about F E1 it .First, the middle category (F E1 it = 3) correponds to the case of the expectation matching with the subsequent perceived realization.Second, the forecast error is allowed to be two-sided, i.e. it can obtain a di¤erent or the same sign on the two sides of the middle category. 9We make use of this property when we study whether households'borrowing behaviour reacts symmetrically or asymmetrically to the optimistic and pessimistic forecast errors.
Our second measure, denoted F E2 it , is asymmetric and has four categories.
Besides allowing for separate categories for those making pessimistic or no forecast errors, this measure distinguishes between prudent and non-prudent optimistic errors.The categories of F E2 it are pessimistic forecast error ( and non-prudentially optimistic forecast error (F It is important to be clear what prudent and non-prudent optimistic forecast errors mean.The prudent optimistic forecast error refers to those cases in which a household experiences a negative surprise relative to what it expected but, despite of that happening, its …nancial situation does not actually get worse. A household is said to experience a non-prudent optimistic forecast error if it both experiences a negative surprise relative to what it expected and if its …nancial situation actually gets worse.The di¤erence between the prudent and non-prudent errors is that the latter is more important for, and likely to call for a greater response by, …nancially fragile households (Lusardi, Scheiner and Tufano 2011).In particular, if the …nancial situation of a household does not actually get worse despite it having made a forecast error, there is less need for adjustment.
Table 2 consists  Reading Panel A, we can see that F E1 it classi…es 4.9% of the households as having made clearly optimistic forecast errors and 17.3% as having made moderately optimistic forecast errors.The corresponding numbers for the pessimistic errors are very close to these, 3.7% and 17.1%, respectively.This con…rms that on average, households'forecast errors have been quite symmetric over the sample period.The numbers for F E2 it , in turn, show that 13.0% of the households appear to have made non-prudentially optimistic errors.These households both experienced a negative surprise relative to what they expected and their …nancial situation actually got worse.The fraction of households that made prudentially optimistic forecast errors is 9.2%.
A particular worry about F E1 it and F E2 it is that they are based on the subjective views of the households and thus that they do not measure anything real.Starting with F E1 it , the second last column in Panel A shows the average change in disposable annual income (from t 1 to t) declines monotonically as we go from the category of clearly pessimistic forecast errors (F E1 it = 1) to the category of clearly optimistic forecast errors (F E1 it = 5).This pattern is what we would expect if households make errors in their expectations and particularly if they have been unable to predict the direction of changes in their disposable income.
The table also shows that the mean drop in disposable income among those households that make the largest, clearly optimistic errors is about around 330 euros.This is lower than the threshold of 1500 euros used by Lusardi, Schneider and Tufano (2011) to measure which European households are likely to be …nancially fragile.
Panel A also shows how shocks in annual disposable income, as measured by a regression residual from an estimated AR(1) process (with year dummies included, using pooled data), vary across the forecast error categories in an expected fashion: the average shock in income becomes monotonically more negative as we go from the category of clearly pessimistic forecast errors to the category of clearly optimistic forecast errors.The last column shows, in particular, that those making largest optimistic errors experience, on average, greatest negative income shocks.For them, the average size of shocks is as high as 4200 euros.
Finally, Panel A shows that also F E2 it works as expected: households making non-prudentially optimistic forecast errors appear to be di¤erent from the rest: Their disposable income develops, on average, clearly less favorably.Their income declines on average about 740 euros, whereas in the other categories, disposable income increases (as it should).Those making non-prudentially optimistic forecast errors experience, on average, largest negative income shocks, as among them, the average size of shocks is about 4500 euros.
[Insert Table 2 about here] The expected (monotonic) patterns documented in Panel A make it clear that F E1 it and F E2 it coincide with real changes in households'…nancial condition.It is of course di¢ cult to take a conclusive stance on whether the households are making errors because of biased ex ante expectations or because they have just experienced truly unexpected …nancial shocks.It is nevertheless interesting to observe that unlike the changes in disposable income, the shocks to annual disposable income, as captured by the regression residual, are on average negative in all forecast error categories.This means that the household-level shocks to disposable income that cannot be predicted using households'lagged income (controlling for year dummies) have on average been negative also for those who made pessimistic forecast errors over the nineteen years that our data cover.
Turning then to Panel B, we …nd that out of those who make clearly optimistic forecast errors, most (87%) make non-prudentially optimistic forecast errors.However, around a third (33%) of those who make non-prudentially optimistic forecast errors make clearly optimistic forecast errors.The two measures are thus not measuring the same thing and are hence complementary.
Finally, Panel C shows that average changes in disposable income are greater in the extreme categories, if we look at the actual outcomes (A it ), as compared to the corresponding categories in the expectations (E itjt 1 ).For example, the disposable income decreases, on average, among those households who expect clearly worse …nancial situation by about 580 euros.On the other hand, the decrease is on average about two thousand euros among the households who report that their …nancial situation actually became clearly worse.
4 Forecast errors and household borrowing behaviour

Household borrowing and indebtedness
The debt measure available from the IDS is comprehensive.It includes many kinds of household borrowing, including mortgages and various forms of nonmortgage debt, such as consumer credit, student loans and loans taken up for the purpose of acquiring income.In our baseline estimations, our measure of debt, D it ; refers to the total amount of loans that household i has accumulated by period t.The debt measure is available for two consecutive periods for each household.
To obtain our main measure for the indebtedness of households we divide D it by the annual disposable income of household i at t: This gives us a debt-toincome ratio, D_I it .This measure captures how much in debt the household is relative to its current disposable income at the end of year t, i.e., after the realization of the forecast error.An important bene…t of this baseline measure is that it allows us to control for the once lagged debt-to-income ratio in the regressions, because each household shows up in our data twice.
The mean of D_I it is 0.55 and its standard deviation is 0.92 (see Table A1 in the Appendix).The median is much lower, because many households have no debt; as we explain below, our estimation results are robust to using a Tobit model, which explicitly allows for clustering at zero.
In our baseline estimations, we regress D_I it on F E1 it and, alternatively, on F E2 it .In all estimations, the omitted caregory refers to those who do not make a forecast error.We consider four di¤erent model speci…cations.First, in Model 1, we regress D_I it on the forecast error categories without any controls.
We report the results of this simple model, because it provides a direct way of showing descriptively how much more (or less) debt those households that make forecast errors have relative to those who do not make such errors.
In Model 2, we control ‡exibly for cohort, age and time e¤ects.To this end, we include long vectors of dummies for the cohort and age e¤ects, which impose only very mild restrictions on the possible lifecycle and cohort patterns in the data. 10As suggested e.g. by Gourinchas and Parker (2002), we model the time e¤ects as a function of underlying macroeconomic and market condition variables. 11The set of variables that we use for this purpose is unusually rich and include, for example, measures for real GDP growth, unemployment, in ‡ation, stock returns and their variability, short interest rates and their variability (see  Model 2 also controls for regional housing prices, which may boost home equity-based borrowing and consumption (Mian and Su… 2011).In addition, it includes dummies for education (5 categories), dummies for the region of residence (20 categories), socioeconomic status of the household (9 categories), type of the household (as determined by the Statistics Finland; 5 categories), size of the household (3 categories), marital status (of the head of the household; 4 categories), and dummies for the gender of household head and whether he/she has recently retired.In addition, Model 2 includes a dummy for the spoken language, which is one for swedish-speaking minority, and a dummy for homeowners (owner-occupiers).Because the dummy for homeowners may not be predetermined, we have checked that none of our estimation results depends on us including it in the model.
Model 3 augments Model 2 by additionally including the lagged disposable income of household i.Given that the dependent variable is the debt-to-income ratio, the lagged disposable income controls for the possibility that making forecast errors is releted to the level of income.
Finally, Model 4 includes the lagged dependent variable as a control variable in addition to all the controls that are already included in Model 3. In these estimations, the lagged dependent variable controls for the stock of accumulated debt.Four main …ndings stand out from Panel A: First, Model 1 shows that when nothing is controlled for, households that make clearly and moderately pessimistic forecast errors have respectively 0.136 and 0.117 euros more debt for each earned euro of disposable income than those who do not make forecast errors.Households that make moderately and clearly optimistic forecast errors also have greater debt levels than those who do not make forecast errors: They have respectively 0.172 and 0.311 euros more debt for each earned euro of disposable income.
Second, as Models 2-4 show, there is a clear asymmetry: The …nding of pessimistic forecast errors being associated with greater indebtedness no longer holds when controls are introduced.However, the result that the households who make optimistic forecast errors are more indebted than households who make no errors is robust across the four models.The F-tests reported in the bottom of the panel show, moreover, that we reject the null hypothesis that the coe¢ cients of clearly and moderately optimistic forecast errors are equal to those of pessimistic forecast errors.This suggests that the households that make clearly and moderately optimistic forecast errors have higher debt-to-income ratios than those who make similar pessimistic forecast errors.
Third, the households that make clearly optimistic forecast errors have higher debt-to-income ratios than those who make moderately optimistic errors.This …nding shows that the magnitude of optimistic errors matters.It also squares with the results of Puri and Robinson (2007), who …nd that extreme, as opposed to moderate, optimism covaries positively with non-prudent economic behaviour.
Fourth, the economic magnitudes are not negligible.For example, as the estimates for Model 4 show, those making clearly optimistic forecast errors accumulate 0.06 euros more debt per each earned euro than those who do not make forecast errors.This e¤ect is not small, when compared to the mean of 0.55 of the outcome variable.
Panel B shows that those making non-prudentially optimistic forecast errors accumulate more debt per each earned euro.Model 4 shows, in particular, that once the past level of indebtedness is held constant, the households making nonprudentially optimistic errors are more indebted than others and also relative to those who make prudentially optimistic errors.
[Insert Table 3 about here] The results reported in

Overindebtedness
We now take a look at how forecast errors covary with the overindebtedness of households.In particular, if the level of indebtedness of the households that make the largest optimistic forecast errors is unsustainable, we expect a positive relation between making such errors and perceived overindebtedness.with the twist that now the three models also include the lagged dependent variable as a control.
[Insert Table 4 about here] The results reported in Panels A and C show that the households that make clearly optimistic forecast errors are more likely than others to report that they are overindebted.There is also evidence that households that make moderately optimistic forecast errors are more likely to report that they are overindebted than those who do not make forecast errors.Panels B and D show that the same holds for the households that make non-prudentially optimistic forecast errors, but not for those who make prudentially optimistic forecast errors.
The economic magnitudes are signi…cant.As Panel C shows, the probability of a household reporting that it has too much debt is about 10% higher for those that make clearly optimistic forecast errors than for those who do not make forecast errors.This is a large di¤erence, considering that the mean of the dependent variable is 5%.From Panel D we see that the corresponding di¤erence is about 8% for those making non-prudentially optimistic forecast errors.
In sum, the results support the view that making large and/or non-prudent optimistic errors covaries positively with unsustainable levels of debt.We do not …nd similar evidence for pessimistic forecast errors.
To explore the robustness of the …ndings reported in Table 4, we have reestimated the models in three ways.First, we replaced the lagged disposable income in Model 3 of each panel by a measure of shocks to disposable income (obtained as the regression residual from a pooled estimation of AR(1) income process, with year dummies included).The results do not change.Second, we included our proxy for the …nancial sophistication index (Calvet, Campbell,   and Sodini 2009) to the most complete model (Model 3) of the panels.To our surprise, it obtained a positive and signi…cant coe¢ cient.Taken literally, this suggests that higher …nancial sophistication is associated with greater likelihood of households reporting that they are overindebted.This sounds somewhat peculiar and casts doubt on us being able to replicate the index properly.Be that as it may, the inclusion of this index did not change our key …nding of optimistic forecast errors being linked to reported overindebtedness.Third, Logit estimations produce very similar results as the linear probability models reported in Table 4.
Table 5 reports results for four alternative indicators of overindebtedness and households experiencing problems with payments: First, we use as the dependent variable an indicator that is one if a household, which has debt, reports in the survey component of the IDS that it has had to agree with its bank to reschedule its debt payments, and that is zero otherwise ("Reschedule"; mean = 0.09).Second, we use an indicator that is one if a household reports that it has had problems in paying its regular (utility etc.) bills ("Bills"; mean = 0.04).Third, we de…ne an indicator that is one if a household, which has debt, reports that it has had problems in meeting its regular debt payments ("Debt payment problems"; mean = 0.07).Finally, the fourth alternative dependent variable is an indicator that is equal to one if a household reports that, given its income from all sources, it has had problems in covering its basic expenditures ("Problems in making ends meet"; mean = 0.07).
In Panels A and B of Table 5, we report the results for F E1 it and F E2 it , respectively, using Model 3 of Table 4.In Panels C and D, we repeat the estimations but so that the lagged dependent variable is included in the models.The size of the estimating sample varies across the columns, because the alternative measures are not available for the same years and/or households.
As can be seen from the table, the results are robust across the panels and columns.Taken together, they support the view that making clearly optimistic forecast errors covaries positively with problems in making payments and ends meet.As Panels C and D of the table show, controlling for the past problems does not change this conclusion.These results also support the view that there is a clear asymmetry: Making pessimistic forecast errors does not covary positively with payment problems.
[Insert Table 5 about here] To explore the robustness of the …ndings reported in Table 5, we implemented the same robustness tests that we did for Table 4 (i.e., replaced the lagged disposable income with the measure of shocks to disposable income; included the proxy of the …nancial sophistication index; and used Logit estimation).The results of Table 5 did not change.

Conclusions
This paper reports three key …ndings: First, households that make the most optimistic forecast errors accumulate greater levels of debt relative to their disposable income.Second, such households are more likely to report that they are overindebted and that they …nd it di¢ cult to cope with their bills and other payments.Third, there is a notable asymmetry, as similar systematic patterns cannot be found for the households that make pessimistic forecast errors.We obtained these …ndings by direct measurement of forecast errors at the level of households and by linking each household's forecast error to its borrowing behaviour and indebtedness, using the same data source.
These …ndings are novel, but in line with Puri and Robinson (2007), who show that extreme optimism covaries positively with non-prudent economic behaviour.Our …ndings are also consistent with Souleles' (2004) results on how optimistic forecast errors are associated with reduced consumption in the US.
We cannot take a conclusive stance on why optimistic forecast errors are positively associated with greater levels of debt and overindebtedness.It is possible that the households making such errors have made poor borrowing choices based on biased ex ante expectations.It is also possible that they have experienced unexpected …nancial shocks which they attempt to smooth by borrowing more.
Be that as it may, it seems that the households that make the most optimistic forecast errors are at risk of accumulating greater and potentially unsustainable levels of debt.Such households may already be or may easily become …nancially fragile.The reason for this is the following: The mean drop in income among those who make the most optimistic forecast errors is lower than the threshold of 1500 euros used by Lusardi This table reports the cross-tabulation of households' expectations (E it|t-1 ) on the development of their financial situation over the next 12 months (year t) and the actual outcomes (A it ), as assessed a year later.The categories take on values "1 = clearly towards better", "2 = somewhat towards better", "3 = stays/stayed about the same", "4 = somewhat towards worse", and "5 = clearly towards worse".The cell entries are the number of households and the corresponding cell probabilities.The data are from Statistics Finland's Income Distribution Statistics (1994-2013).
Actual outcome (A it ) Obs.

Percent
Cum.

Change in income
Shocks to income This table reports the cross-tabulation of households' expectations (E it|t-1 ) on the development of their financial situation over the next 12 months (year t) and the actual outcomes (A it ), as assessed a year later.The categories take on values "1 = clearly towards better", "2 = somewhat towards better", "3 = stays/stayed about the same", "4 = somewhat towards worse", and "5 = clearly towards worse".The cell entries are the number of households and the corresponding cell probabilities.The data are from Statistics Finland's Income Distribution Statistics (1994-2013).
Actual outcome (A it ) Obs.

Percent
Cum.

Change in income
Shocks haviour.The second component of the IDS data comes from an interview-based database of Statistics Finland, called Income and Living Conditions Survey.This survey component includes, for example, questions about the expectations of households about the subsequent development of their …nancial situation.
of three panels.Panel A displays the distribution of the data by F E1 it and F E2 it and the average change in disposable income, measured in two alternative ways, for each forecast error category of the two measures.Panel B displays the cross-tabulation of F E1 it and F E2 it .Finally, Panel C reports the average change in disposable income by outcome, A it , and expectation, E itjt 1 , categories.

Table A1
in the Appendix).

Table 3
reports the OLS results for Models 1-4: Panel A gives the results when F E1 it is used and Panel B for F E2 it .The standard errors are clustered at region-year level.The lower part of the panels reports F-tests for the di¤erences in the estimated coe¢ cients of F E1 it and F E2 it across the various forecast error categories.

Table 3
are robust to a number of alternative speci-…cations (see Online Appendix for details).For example, the results hold when we use the Tobit model to account for the cluster of zeros in the dependent variable; if we use the debt-payment-to-income ratio (i.e., mortgage-service to income -ratio) as an alternative outcome measure; if we decompose D_I it into its mortgage debt and non-mortgage debt components; if we replace the lagged disposable income by a measure of shocks to disposable income; and if we add a proxy for households'…nancial sophistication(Calvet, Campbell, and Sodini   2009)to the estimated models.
, there is no agreement in the literature on what the overindebtedness of households means or how it should be measured.The standard model of consumption and intertemporal allocation (see, e.g., Attanasio and Weber 2010) predicts that households attempt to smooth their consumption intertemporally and that borrowing (or deleveraging) is an important part of such smoothing (Sullivan 2008, Hurst and Sta¤ord 2004).The possibility that households could carry suboptimal levels of debt is rarely explicitly considered (see however Lusardi and Tufano 2015 and Gathergood 2012).Due to lack of an established practise, we use a number of indicators of overindebtedness (and problems in making payments).Our primary indicator is a dummy that is one if a household reports in the survey component of the IDS the perception that it has too much debt, and is zero otherwise.The measure is similar in spirit to that used byLusardi and Tufano (2015).In our data, its mean is 0.05 (see TableA1in the Appendix).The measure is not available over all the sample years.The estimating sample is therefore smaller in what follows.
Table4reports the estimates from linear probability models in which the dummy for the perceived overindebtedness is the dependent variable.There are four panels in the table, each of which reports the results from three model speci…cations: Panel A reports the results for F E1 it and Panel B for F E2 it .The R.H.S. speci…cations of the three columns in these panels match with Models 1-3 that we used for Table3.That is, in Model 1, we have no controls.Model 2 controls ‡exibly for the cohort, age and time e¤ects and also includes the long vector of demographic controls.Model 3 additionally includes the lagged disposable income.Panel C and D repeat the analyses of Panel A and B, but

Table 1 :
, Schneider and Tufano (2011) to measure which European households are …nancially fragile.And yet we …nd that those who make clearly or non-prudent optimistic forecast errors are more likely to report being overindebted and having problems in making ends meet.This suggests that making optimistic forecast errors and being …nancially fragile are connected.Cross-tabulation of expectations and actual outcomes Moreover, in an earlier version of our paper, we explored how the forecast errors correlate with the observed characteristics of the households.It seemed that some of the variables that explain variation in the likelihood of making forecast errors, such as low education, are those that according to the prior literature are related to poor …nancial literacy (Calvet, Campbell and Sodini 2009; Bruine de Bruin et al. 2010).This suggests an interesting avenue for further research, as poor …nancial and economic literacy may explain why those who make forecast errors accumulate more and perhaps unsustainable levels of debt.There already is some cross-sectional evidence pointing to this direction, as it seems that overindebted households in the UK are …nancially less literate and more likely to have experienced various economic and …nancial shocks(Gathergood 2012).

Table 2 :
Distribution and cross-tabulation of forecast errorsPanel A reports the distribution of the data by the forecast error categories and average changes in disposable income for each category.The average change in disposable income is measured in two alternative ways: by the annual change in disposable income and by the shocks to income, as measured by the residual of the regression of income on lagged income and year dummies.Panel B reports the cross-tabulation of the two forecast error measures (FE1 it and FE2 it ).

Table 2 ,
continued: Panel B: Cross-tabulation of forecast error measures (FE1 it and FE2 it )

Table 3 :
Forecast errors and indebtednessPanels A and B report OLS estimates of the effect of forecast errors (FE1 it and FE2 it , respectively) on the debt-to-income ratio.Standard errors are clustered at the region-year level and reported in parentheses; *** p<0.001, ** p<0.01, * p<0.05.The data are from the Statistics Finland's Income DistributionStatistics (1994-2013).

Table 4 :
Forecast errors and overindebtednessPanels A and B report OLS estimates of the effect of forecast errors (FE1 it and FE2 it , respectively) on perceived overindebtedness.Panels C and D report the same models including a lagged dependent variable as a control.Standard errors are clustered at the region-year level and reported in parentheses; *** p<0.001, ** p<0.01, * p<0.05.The data are from the Statistics Finland's Income DistributionStatistics (1996-2013).

Table 5 :
Forecast errors and alternative indicators of overindebtednessPanels A and B report OLS estimates of the effect of forecast errors (FE1 it and FE2 it ,

Table 1 :
Cross-tabulation of expectations and actual outcomes

Table 2 :
Distribution and cross-tabulation of forecast errorsPanel A reports the distribution of the data by the forecast error categories and average changes in disposable income for each category.The average change in disposable income is measured in two alternative ways: by the annual change in disposable income and by the shocks to income, as measured by the residual of the regression of income on lagged income and year dummies.Panel B reports the cross-tabulation of the two forecast error measures (FE1 it and FE2 it ).

Table 2 ,
continued: Panel B: Cross-tabulation of forecast error measures (FE1 it and FE2 it )

Table 3 :
Forecast errors and indebtednessPanels A and B report OLS estimates of the effect of forecast errors (FE1 it and FE2 it , respectively) on the debt-to-income ratio.Standard errors are clustered at the region-year level and reported in parentheses; *** p<0.001, ** p<0.01, * p<0.05.The data are from the Statistics Finland's Income DistributionStatistics (1994-2013).

Table 4 :
Forecast errors and overindebtednessPanels A and B report OLS estimates of the effect of forecast errors (FE1 it and FE2 it , respectively) on perceived overindebtedness.Panels C and D report the same models including a lagged dependent variable as a control.Standard errors are clustered at the region-year level and reported in parentheses; *** p<0.001, ** p<0.01, * p<0.05.The data are from the Statistics Finland's Income DistributionStatistics (1996-2013).

Table 5 :
Forecast errors and alternative indicators of overindebtednessPanels A and B report OLS estimates of the effect of forecast errors (FE1 it and FE2 it , respectively) on alternative indicators of overindebtedness.Dependent variables in models 1-4, respectively, are dummy variables as follows: having new loan repayment schedule ("Reschedule" = 1), having problems in paying regular bills ("Bills" = 1), having problems in meeting regular debt payments ("Debt payment problems" = 1), and having had problems in making ends meet ("Problems in making ends meet" = 1).Panels C and D report the same models including a lagged dependent variable as a control.Standard errors are clustered at the region-year level and reported in parentheses; *** p<0.001, ** p<0.01, * p<0.05.The data are from the Statistics Finland's Income Distribution Statistics (different data periods).