A good time to buy: Predicting Dutch household consumption with consumer sentiment

In this paper we argue that the sentiment sub-index measuring whether now is “a good time for large purchases” is more informative than the headline measure. That sub-index fell by 11 index points recently, implying a reduction in monthly household consumption of approximately 0.11 percentage points. Statistics Netherlands publishes a monthly composite consumer confidence index along with several sub-indices that capture different dimensions of consumer sentiment. These survey-based measures provide unique insights into consumers’ feelings about current and future economic conditions rather than derived measures based on purchases. Consumers answer questions about their sentiments regarding past, current, and future economic conditions. The underlying behavioral mechanism implied is that when households feel confident about their current and future economic situation, their discretionary spending may increase (Katona, G. 1968; European Commission, 2026). In this paper, we take the survey responses at face value, assuming they accurately reflect the sentiments of respondents at the time of the survey.

Yet the relationship between confidence and consumption is empirically fragile. The headline consumer confidence index has a mixed forecasting record, and it is not clear which aspect of sentiment, expectations about the broader economy (assessments of personal finances, or willingness to make purchases) drives the connection, if it exists at all (Ludvigson, 2004; Carroll et al., 1994). Fundamental differences across countries in terms of economic safety nets, export dependence, and positions within the international economy suggest that the effect of consumer sentiment on consumption might be country-specific.

In this paper we investigate whether consumer confidence, as measured by Statistics Netherlands, influences household consumption patterns in the Netherlands, and which measure is the strongest predictor of changes in household consumption. We do so by using a Bayesian Vector Autoregression (BVAR) framework with a Minnesota prior, estimated on monthly data from 2003 to the present (Lütkepohl, 2005). By introducing shocks into a well-specified model, we can identify the extent and duration of changes in consumption correlated with changes in confidence. To identify which confidence measure matters most, we estimate separate models for each forward-looking indicator and compare impulse response functions using a common scale, allowing for direct comparison of the strength of each confidence measure in predicting household consumption.

Which confidence measure best predicts consumption?

We estimate a separate BVAR for each of the forward-looking Statistics Netherlands confidence measures and compare the response of consumption to a ten index-point drop in confidence. We exclude three backward-looking indices from the analysis: economic climate (a composite measure that combines backward- and forward-looking components), economic situation over the past 12 months, and financial situation over the past 12 months. Models estimated using these measures produce either flat or inverted impulse responses, consistent with the theoretical expectation that retrospective assessments are outcomes of the consumption process rather than drivers of it. The fact that the modeling framework consistently picked up differences between forward- and backward-facing indicators provides some assurance that the measures capture the concepts they are intended to measure and that the model behaves as expected.

Figure 1 shows the consumption response to a 10 index-point drop in consumer confidence. The solid orange line shows the response of the “a good time for large purchases” measure. The shaded areas represent the 94% highest density interval (HDI) around the median posterior response.

The figure plots the rescaled IRFs for all four forward-looking measures. We report additional robustness results in table 1. Both the figure and the table reveal a clear distinction in terms of the strength of the measures. The sub-index measuring whether the current moment is a good time for large purchases produces the strongest and most precisely estimated consumption response. The peak effect is 0.10 percentage points and occurs at horizon 0; in short, the response is contemporaneous. The corresponding 94% credible interval is borderline in that it nearly excludes zero, although most of the distribution lies far from zero. This result implies that a 10 index-point drop in this index leads to a 0.10 percentage point drop in monthly household consumption. Statistics Netherlands data for April 2026 shows that the “a good time for large purchases” sub-index declined by 11 points from the previous month, which, given linearity, means a monthly drop in consumption of 0.11 percentage points. The median responses for the other consumer measures also show negative, yet weaker responses. However, their credible intervals include zero, suggesting we cannot rule out zero effect on consumption.

The headline consumer confidence index broadly mirrors the pattern of the other measures. This is unsurprising, given that it combines both forward-looking and backward-looking sub-indices; more precisely, the consumer confidence index is calculated as the simple average of all five measures used by CBS to assess overall consumption. Since the backward-looking sub-indices perform poorly when predicting current and future consumption, their inclusion likely weakens the predictive power of this measure. The relatively poor performance of the series “expected economic situation in the next 12 months” and “expected personal financial situation over the next 12 months” is more surprising, given that they are forward-looking. One possible explanation is that “a good time for large purchases” asks a much more direct question about consumers’ immediate willingness to spend, rather than measuring their means or ability to make purchases in the future.

Why buying intentions matter

One possible explanation for this ranking lies in the psychology of consumer decision-making. “A good time for large purchases” measures whether households believe that now is a good moment to make large purchases, making it the sub-index most directly connected to actual spending decisions. The question is intended to identify large durable purchases, which have been found to have a positive and statistically significant effect on expenditures (Mynaříková & Pošta, 2023). By contrast, expectations about the macroeconomy or personal finances over the coming year are perhaps more speculative and less directly connected to the immediate decision to spend. These findings suggest that the Statistics Netherlands measure “a good time for large purchases” may be a timely and reliable early indicator of durable household spending. Because it is published monthly and precedes official consumption data by several weeks, policymakers and businesses can use it to gauge near-term spending developments.

Data

We use monthly data for the Netherlands covering the period 2003-2026. The starting date is constrained by the availability of the employment series, which begins on February 1, 2003. Each model includes four time series (explained below): the DNB mortgage interest rate; the unemployment rate (ages 15 to 74, seasonally adjusted), household consumption (volume index, 2021=100), and one of four forward-looking consumer confidence measures – all sourced from Statistics Netherlands. We select the series following Dées and Soares Brinca (2013), with some adjustments to take into account features of the Dutch economic system. We select the interest rate to capture monetary conditions and the cost of credit, both of which directly affect the affordability of large purchases. We tested several interest rates and ultimately selected the Dutch 10-year government bond yield. Because relevant interest rates are often cointegrated, the choice of which specific rate is somewhat arbitrary. The selected rate primarily serves to control for the effects of monetary policy on the relationship between confidence and consumption.

Similarly, the unemployment rate is included to control for labor market effects in the relationship between confidence and consumption. These variables capture labor market conditions and income uncertainty, which are key drivers of precautionary saving behavior under the Permanent Income Hypothesis (Friedman, 1957). Together these two variables control for the macroeconomic environment within which confidence is formed, allowing us to isolate the independent contribution of sentiment to consumption.

We test the following confidence measures: the composite indicator consumer confidence (consumentenvertrouwen) and three sub-indices: “a good time for large purchases” (gunstige tijd voor grote aankopen), “expected economic situation over the next 12 months” (economische situatie komende 12 maanden), and “expected personal financial situation over the next 12 months” (financiële situatie komende 12 maanden). Figure 2 plots the measures alongside household consumption over the sample period. While differences in volatility make it difficult to identify short-term patterns, longer-term (annual) correlation patterns are more apparent.

We are fortunate that Statistics Netherlands publishes monthly household consumption data, allowing us to avoid the information loss or artificial smoothing that can result from interpolating quarterly national accounts data. We seasonally adjust consumption and convert it to month-on-month percentage changes. The interest rate and unemployment rate are both expressed in percentage terms. All confidence indices enter the model in levels, following standard practice (Dées & Soares Brinca, 2013; Sims, 1980). Seasonal adjustment of consumption was not entirely successful. We partially addressed a remaining 12-month autocorrelation pattern by including additional lags and choosing a less restrictive prior to absorb some of the autocorrelation.

Model specification

Using these series, we estimate BVAR models with six lags, with one model per confidence measure. We prefer the BVAR over a standard VAR for several reasons. First, it generates a joint posterior multivariate probability distribution from which we can, through a series of steps, draw samples to construct credibility intervals. Second, it allows regularization through a Minnesota prior which shrinks the number of lag coefficients toward zero (Litterman, 1986; Doan, Litterman and Sims 1984). The BVAR framework allows us to include more series with a greater number of lags, thereby mitigating potential omitted variable biases (Giannone, Reichlin and Small, 2006).

The Minnesota prior starts from the assumption that each series follows a random walk, for example, the best forecast of next quarter’s consumption is consumption this quarter. The Bayesian approach is conservative in that it allows the data to pull coefficients away from their starting priors via the likelihood, provided that the data-derived evidence is strong enough. This feature is particularly important in our setting: with four variables, six lags, and 277 observations, an unrestricted VAR would estimate 24 lag coefficients per equation, risking overfitting. The prior and modeling framework imposes discipline without discarding the data.

Zivot-Andrews unit root tests, which allow for structural breaks in a series, fail to reject the null hypothesis of a unit root for the interest rate, unemployment, and confidence sub-indices. Household consumption, expressed as a month-on-month percentage change, rejects the null and is therefore considered stationary by construction. Rather than differencing the remaining variables and thereby rendering those stationary, we again rely on Minnesota priors and follow the standard approach in the BVAR literature and estimate the model in levels. Priors mitigate any remaining non-stationarity by preventing explosive dynamics even in the presence of near-unit-root behavior (Sims, 1980; Bańbura, Giannone and Reichlin, 2010). Differencing, while a common approach to addressing non-stationarity, would remove potentially informative long-run level information. For instance, whether the mortgage rate is at 2% or 5% matters for consumption decisions in a way that the month-on-month change cannot capture.

Shock identification

We identify shocks using a Cholesky decomposition, with variables ordered from most to least exogenous: interest rate, unemployment, consumer sentiment, and consumption. This ordering determines which series affect others contemporaneously, as well as the direction of that influence; for subsequent periods series interactions are unconstrained. For example, in our model no other series affect rate contemporaneously, except its own shock, while the other series can be affected by rate contemporaneously. The economic reasoning is that rates are taken as given at any moment and affect the other series, while those series don’t affect rates. Similarly, consumer sentiment is not allowed to affect unemployment contemporaneously, although it can respond to labor market conditions. Similarly, at a given moment, consumer confidence can affect consumption, while consumption is not allowed to affect consumer sentiment. The ordering used follows Dées and Soares Brinca (2013) and is standard in the consumption-confidence literature (Ludvigson (2004); Carroll et al. (1994)).

We use the estimated model to simulate how an economic shock to one series propagates through the other series in the system over time. These simulated paths, known as impulse response functions (IRFs), trace the dynamic response of each series to a one-standard-deviation structural innovation in each of the other series. To objectively compare impulse response functions across measures, we rescale the units in which each IRF is expressed to a common unit.

This is necessary because the confidence measures have different standard deviations, e.g., a standardized shock to “a good time for large purchases” index is not the same magnitude as a standardized shock to the headline consumer confidence measure, making direct comparison misleading.

Covid-19 adjustment

The Covid-19 pandemic has a substantial distorting effect on models and the reliability of diagnostic tests. Using a Chow test, we confirm a statistically significant structural break in March 2020 and during the period thereafter. The nature of the break matters for interpretation: The sharp collapse and subsequent recovery in consumption during 2020 presumably reflects lockdown measures and reopening dynamics, rather than confidence-driven behavior. To limit the influence of the most extreme observations, we introduce two Covid-19 pulse dummies for March and April 2020. Omitting these dummies inflates the estimated IRF effects by a factor of three, consistent with the pandemic having simultaneously caused a sharp drop in confidence and a collapse in consumption. In addition, we ran all the models for the full period, the pre-Covid period, and the post-Covid period. We then compared the results using a range of statistical tests. The results for “a good time for large purchases” were generally robust over the three periods, although the magnitude and statistical certainty varied across periods.

Robustness checks

As a further check on the influence of the Covid-19 pandemic, we re-estimate each model on the pre-pandemic sample (2003-02 to 2020-02) as shown in table 1. The impulse responses are nearly identical to those obtained from the full sample, confirming that the combination of the Minnesota prior and the Covid dummies is sufficient to prevent the pandemic period from distorting the estimated confidence-consumption relationship. Multivariate portmanteau tests (Hosking, 1980; Li & McLeod, 1981) indicate that the model is well-specified over the pre-pandemic sample (p=0.101). By contrast, the full-sample rejection is driven by the structural break in March 2020, which the Covid dummies are intended to absorb.

To complement the Bayesian estimation, we test Granger causality using a frequentist VAR framework. While the Cholesky ordering sets contemporaneous directional effects, Granger causality tests provide information about the flow of information between two or more series in a model. The IRF identifies the magnitude and duration of, for example, the consumer confidence-consumption relationship, while Granger causality tests address the narrower question of whether consumer confidence has predictive power for consumption beyond the history of consumption itself. As shown in table 1, pairwise Granger causality tests confirm that “a good time for large purchases” is the only sub-index that significantly predicts the consumption in both the full sample and the pre-Covid sample (p=0.023 and p=0.020, respectively). The other forward-looking measures fail to predict consumption once past consumption is taken into account. Importantly, the reverse relationship (consumption predicting confidence) is significant in the full sample for several measures but disappears entirely in the pre-Covid sample for all measures, ruling out reverse causality as an explanation for the “a good time for large purchases” result.

A lag sensitivity analysis, which tests whether the Granger causality result holds across different lag specifications, further supports this finding. In the pre-Covid sample, “a good time for large purchases” uniquely and robustly rejects the null hypothesis of no Granger causality, beginning from lag 9 onward (p≈0.001-0.002). The other forward-looking sub-indices only appear significant at lags 12-18, where degrees of freedom are thin and overfitting is a concern. VAR-based block-exogeneity tests, which account for the joint dynamics of all variables in the system, confirm these findings.

Several limitations should be noted. First, the model assumes a linear and symmetric relationship between confidence and consumption, meaning that a drop in “a good time for large purchases” is assumed to have the same estimated effect, as an equivalent rise. Second, the Covid-19 dummies are a pragmatic solution to account for a major structural break rather than a structural representation of pandemic-related economic dynamics. Finally, the results are specific to the Netherlands and reflect the country’s particular social and institutional features, which may not be applicable to other countries.

Several directions for future research warrant further investigation. The backward-looking sub-indices and the reverse causality channel in which consumption influences confidence deserve separate study, as they may provide useful indicators of current economic conditions rather than predictors of future spending. The European Central Bank’s (ECB) Consumer Expectations Survey offers a promising alternative set of measures. However, it is currently limited to the post-2020 period, making reliable estimation difficult. Revisiting this relationship as the sample grows would be worthwhile. Finally, extending the analysis to other small open economies with similar institutional features, such as Belgium, Austria, or the Nordic countries, might help determine whether the “a good time for large purchases” result reflects a universal feature of consumer behavior or a phenomenon specific to the Netherlands.

We compare five Statistics Netherlands consumer confidence measures, comprising three forward-looking and two backward-looking sub-indices. Only the three forward-looking sub-indices showed evidence of helping to forecast consumption. While consumer confidence has declined sharply in recent months, that measure alone provides limited information about how Dutch household consumption will respond. Because it combines both forward- and backward-looking components, the headline index is not strongly associated with consumption. Our results show that the sub-index, “a good time for large purchases,” helps to forecast consumption, implying that the 11-point decrease recorded in April 2026 will result in a 0.11 percentage point decline in monthly household consumption growth. The finding is robust across both the full sample and the pre-Covid sample.

See also the Dutch-language, non-technical version of this analysis:

Van vertrouwen naar bestedingen: de rol van 'gunstige tijd voor grote aankopen' - Rabobank

Bańbura, M., Giannone, D., & Reichlin, L. (2010). Large Bayesian vector auto regressions. Journal of Applied Econometrics, 25(1), 71-92.

Carroll, C. D., Fuhrer, J. C., & Wilcox, D. W. (1994). Does consumer sentiment forecast household spending? If so, why? American Economic Review, 84(5), 1397-1408.

Dées, S., & Soares Brinca, P. (2013). Consumer confidence as a predictor of consumption spending. Economic Modelling, 31, 15-22.

Doan, T., Litterman, R. B., & Sims, C. A. (1984). Forecasting and conditional projection using realistic prior distributions. Econometric Reviews, 3, 1-100.

European Commission (2026). Methodological user guide: The joint harmonised EU programme of business and consumer surveys, DG ECFIN, Brussels.

Friedman, M. (1957). A theory of the consumption function. Princeton University Press.

Giannone, D., Reichlin, L., & Small, D. H. (2006). Nowcasting GDP and inflation: The real-time informational content of macroeconomic data releases (ECB Working Paper No. 633). European Central Bank.

Katona, G. (1968). On the function of behavioral theory and behavioral research in economics. American Economic Review, 58, 146-149.

Litterman, R. B. (1986). Forecasting with Bayesian vector autoregressions — Five years of experience. Journal of Business & Economic Statistics, 4(1), 25-38.

Ludvigson, S. C. (2004). Consumer confidence and consumer spending. Journal of Economic Perspectives, 18(2), 29-50.Lütkepohl, H. (2005). New introduction to multiple time series analysis. Springer Science & Business Media.

Mynaříková, L., & Pošta, V. (2023). The effect of consumer confidence and subjective well-being on consumers’ spending behavior. Journal of Happiness Studies, 24, 429-453. https://doi.org/10.1007/s10902-022-00603-5.

Sims, C. A. (1980). Macroeconomics and reality. Econometrica, 48(1), 1-48.