2.1. The nexus between confidence and the macroeconomy

The economic studies on the psychological factor can be traced back to Jevons W S [19]. He discussed humans’ pleasure and pain and, thus, the measurement of feelings and motives. Based on this, he proposed the theory of utility called “marginal revolution,” in which psychological factors were introduced into economic analysis through utility functions when studying the determinants of demand. This is an essential start for understanding how psychological factors influence economic decisions and behaviour. After the Great Depression, Keynes J M [20] argued that expectation is a crucial determinant of output and employment. He differentiated short-term and long-term expectations and discussed how three critical psychological factors, the propensity to consume, liquidity preference, and expectations of future returns on capital, affect the total demand. This is the so-called theory of effective demand, in which expectation is crucial for individuals and firms. The work of Keynes J M [20] is one of the bases of modern economics. In short, Jevons W S [19] and Keynes J M [20] underscore the importance of human expectations to the economy as never before.

Although rational expectation (Muth [21]) has been widely used in modern economic research, with the development of behavioral economics, neuroeconomics, and neuroscience, the theoretical assumptions proposed by rational expectation, such as no systematic cognitive bias, consistent beliefs, and sufficient information, have been critiqued by some studies, such as Akerlof et al. [22] and Shiller [23]. Bounded rationality (Simon [24]) and imperfect rational expectations of information (Angeletos [25]) have gradually attracted academic attention. These studies have contributed significantly to understanding how psychological factors affect economic behaviour and the macroeconomy.

In recent years, research on confidence in the economy has mainly focused on economic growth (business cycle), fiscal and monetary policy, inflation, and prices.

For the nexus between confidence and economic growth (business cycle), existing studies have shown that consumer and entrepreneurial confidence are procyclical (Taylor and McNabb [7]). Consumer and entrepreneurial confidence improvements can stimulate domestic demand and promote economic growth (Liu et al. [26]). In other words, confidence positively correlates with economic growth (Guo and He [10]). As a result, many studies have incorporated confidence in economic growth forecasting and found that it improves the accuracy of predictions (Giannone et al. [27], Jiang and Li [28]). Nevertheless, some studies argue that although entrepreneurial confidence can influence macroeconomic fluctuations, consumer confidence cannot (Chen and Tang [13]). For the impact of confidence shocks on the macroeconomy, Zhuang et al. [9], Zhu and Deng [11], and Dees and Zimic [12] found that both expectation shocks and noise shocks have positive effects on main macroeconomic variables, and expectation shocks can explain a significant portion of economic growth fluctuations.

Fiscal and monetary policy are essential pillars of the macroeconomy in modern economics. How confidence affects the macroeconomy is a meaningful topic. Wang et al. [5] and Zhang and Sun [6] found that confidence plays a significant role in the transmission mechanism of monetary policy. Removing the bridging role of confidence could significantly reduce the effectiveness of monetary policy. Guo and He [10] and Li [29] found that the influence of confidence on the policy effect varies across different economic growth stages. Restoring confidence can amplify the policy effect and thus alter the policy effect multiplier. These results analyze the impact of confidence on a specific macroeconomic field, enriching the research on the relationship between confidence and the macroeconomy.

With the development of the internet, some studies attempt to utilize big data to study confidence. For example, Dong and Bollen [30], Liu and Li [31], and Yuan et al. [32] constructed internet consumer confidence indices with search engine data and explored their relationships with industrial value added, investment, consumption, and price. The results revealed that internet consumer confidence is a Granger cause of industrial value added and can explain approximately 20% of the variation in industrial value added. Changes in confidence also lead to investment, consumption, and price changes.

Recently, a growing body of studies on expectation has been using textual methods to extract the sentiment of media and then explore the effect of sentiment on the macroeconomy. For example, Larsen and Thorsrud [16] and Ardia et al. [33] showed that sentiment indicators computed by retrieving newspaper articles have predictive power for key economic variables, such as industrial production. Thorsrud’s [34] results suggest that the model with a daily business cycle index based on quarterly GDP growth and textual information is competitive with forecast combination systems and expert judgment. The results of a comparison study conducted by Ellingsen et al. [35] suggest that the news data contains information not captured by the hard economic indicators and that the news-based data are particularly informative for forecasting consumption developments.

Admittedly, the predictive power of sentiment indicators based on news articles prevails over survey data. However, it is worth noting that news articles are freely published and mainly convey the attitude of the media and authorities. In other words, the overall representativeness of the news is relatively poor. On the contrary, survey data is based on scientifically designed questionnaires and rigorous sampling survey theories. Thus, it is more representative of the population. Consequently, survey data is a better candidate for testing the nexus between confidence and the macroeconomy rather than forecasting the macroeconomy.

2.2. Decomposition of confidence

Furthermore, some researchers have further focused on the decomposition of confidence to explore the relationship between confidence and the macroeconomy. Harrison and Weder [36] and Barsky and Sims [37] proposed a decomposition of confidence into “fundamental confidence” and “animal spirits”. The former captures the part determined by macroeconomic fundamentals. Market participants estimate it based on macroeconomic information. The latter represents purely psychological or emotional factors that economic fundamentals cannot explain.

In fact, for confidence shocks, the literature has also adopted other terms, such as “sunspots” (Cass and Shell [38]), “self-fulfilling prophecies” (Azariadis [39]), and “waves of optimism and pessimism” (Chauvet and Guo [40]). Although these concepts have different names, they all refer to external uncertainties or shocks that are unrelated to economic fundamentals (Chen and Tang [13], Cass and Shell [38], Chauvet and Guo [40]). Among these concepts, “animal spirits” is the most widely used and well-known, so we adopt it in this paper.

The most popular research on “animal spirits” is conducted by Keynes [20]. He argued that investors make decisions not only based on mathematical expectations but also on irrational “animal spirits”. This assumption was adopted by Shleifer [41] when explaining why economic activities are aggregated. He concluded that people’s economic behaviour is influenced not only by economic factors but also by irrational factors such as emotions and beliefs. In recent years, some studies have shown that pessimistic consumers and entrepreneurs existed before several of the United States’ economic recessions, even when economic fundamentals were strong, suggesting that changes in consumer and investor perceptions are likely to be important factors influencing the business cycle (Dees and Zimic [12], Chauvet and Guo [40]). Moreover, the spread of pessimistic sentiment exacerbates crises and slows economic recovery (Assenza et al. [42]). Based on this, Nobel laureate Robert J. Shiller [23,43] systematically studied the impact of investor irrational sentiment on stock markets, bond markets, and real estate markets. Along with another Nobel laureate, George Akerlof [44], he attributed the concept of “animal spirits” as one of the root causes of the global financial crisis.

Similar studies have been done in China, but the conclusions are different. For example, Sui et al. [45] found that while “sunspot shocks” significantly impact endogenous variables, their role in transmitting fundamental shocks and fluctuating endogenous variables is weak, while Zhuang et al. [9] and Zhu et al. [11] showed that noise shocks have a positive effect on the macroeconomy.

2.3. Discussion about the existing research

In summary, existing studies expand the research on the relationship between macroeconomics and confidence. However, the nexus between macroeconomy and confidence is still controversial. Some studies suggest that the impact of confidence on the macroeconomy (Zhuang et al. [9], Dees and Zimic [12], Zhu and Deng [11]) is significant, and incorporating expectations retrieved from news articles or survey data into models can improve the accuracy of forecasting. Still, some studies provide opposite evidence, especially for China (Chen and Tang [13]).

There may be two reasons for this difference. One is the measurement error and sample size; the other is the limitation of the econometric model. Specifically, due to the late start of surveys on confidence, most early studies either adopted proxy indicators for confidence or used small samples. This leads to problems with sample size and measurement errors. Meanwhile, limited by econometric methods, most studies employed the same-frequency data models, which strictly require common-frequency data. The mixed-frequency variables are either replaced with proxy variables or aggregated through totalling or averaging. This transformation reduces the sample size but also results in the disorder of causal relationships between variables (Ghysels [17]). For instance, when using mixed-frequency and common-frequency data, Yuan and Wen [18] discovered a significant difference in the causal relationship between economic growth and consumer confidence.

Overall, the relationship between confidence and the macroeconomy and whether this relationship is robust in the context of mixed-frequency data needs more evidence, especially for the post-COVID era in China. Thus, this study adopts mixed-frequency data models and the newest sample to test the nexus between confidence and the macroeconomy. Mixed-frequency data methods can expand the sample size as much as possible. With the latest sample, the results can provide the newest and robust evidence of the relationship between confidence and the economy, especially in the context of mixed-frequency data and the post-COVID-19 era in China.

3.1. Variables and data

This study focuses on the relationship between confidence and the macroeconomy. Following Chen and Tang [13], Liu et al. [26], and Zhu and Deng [11], we measure confidence with the consumer confidence index (CCI). We depict the macroeconomy from five aspects: total output, development of the industrial entity, the financial market, price, and the labour market. Specifically, according to Harrison and Weder [36] and Chen and Tang [13], we use the GDP growth rate (GDP), the consumer price index (CPI), the purchasing managers’ index (PMI), the real estate climate index (RECI), the annual loan interest rate (R), the Shanghai stock composite index (Stock), and the urban unemployment rate (UE) to portray the macroeconomy. All monthly data ranges from January 2005 to June 2024. Accordingly, all quarterly data range from the first quarter of 2005 to the second quarter of 2024. The variables’ definitions, frequencies, and sources are summarized in Table 1.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Variable definitions and data sources.

https://doi.org/10.1371/journal.pone.0332909.t001

Fig 2 plots the raw level of these series. Intuitively, the monthly series shows a certain seasonality. Thus, we make seasonal adjustments to them with X-13ARIMA-SEATS. The quarterly GDP is the year-on-year growth rate, so there is no seasonality. The quarterly urban unemployment rate shows no apparent seasonality.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Time series of variables’ raw level.

The X-axis indicates the time (quarter or month). The Y-axis indicates the raw levels of variables. See Table 1 for definitions of variables.

https://doi.org/10.1371/journal.pone.0332909.g002

Lastly, non-stationary data may lead to misleading results in causality tests and RU-MIDAS. Thus, we conduct ADF tests before the formal analysis. The results are displayed in the S1 Table. It is shown that the levels of the consumer confidence index and the real estate market index are . GDP, CPI, the purchasing managers’ index, and the urban unemployment rate have a deterministic trend. Thus, we make a difference (detrending) to the consumer confidence index and the real estate market index (GDP, CPI, the purchasing managers’ index, and urban unemployment rate), and adopt the data after difference and detrending in the following analysis.

3.2. Model specification

As mentioned above, performing temporal aggregation when modelling mixed-frequency data is common, especially in macroeconomic research. Specifically, macroeconomic data are often low-frequency due to statistical and economic costs. Annual and quarterly data are prevalent, while monthly, weekly, and daily data are scarce. Therefore, it is necessary to aggregate or average high-frequency data to low-frequency data in many scenarios. This approach not only leads to the loss of high-frequency information and decreases the sample size but also results in the distortion of causal relationships between variables (Ghysels [17]). To avoid this problem, mixed-frequency data models have been developed and applied rapidly in recent years. They have been widely used in economic and financial research, for example, Cui and Zhang [46] and Lu et al. [47]. The best advantage of these models is that they can incorporate the mixed-frequency data directly and consequently effectively maintain the actual relationships among the variables. These methods have attracted substantial attention because of their ability to utilize the information in unbalanced frequency datasets directly.

For our case, mixed-frequency data, such as the quarterly consumer confidence index and monthly CPI, will be used to test the relationship between confidence and the macroeconomy. Thus, mixed-frequency data models are better candidates. This paper uses mixed-frequency vector autoregression, mixed-frequency Granger causality tests, and reverse-constrained mixed-frequency data sampling to test this relationship. More details about these models can be found in Ghysels [48] and Foroni et al. [49].

3.2.1. Mixed-frequency vector autoregressive models and mixed-frequency Granger causality test.

There are currently two forms of mixed-frequency vector autoregressive models (MF-VAR). The first form assumes that all variables in the mixed-frequency economic system share common high-frequency data-generating processes, while the low-frequency variables are latent and missing periodically. Thus, state space models can be used for modelling and analysis. For more details, see Mariano and Murasawa [50] and Lu et al. [47]. However, the parameter estimated by this approach cannot be directly used for the Granger causality test (Ghysels et al. [17], Yuan and Wen [18]). Therefore, Ghysels et al. [17] proposed another form. This approach is more appropriate for our case.

Let and denote high- and low-frequency stationary variables, respectively. The ratio of sampling frequencies is . In this paper, and are monthly and quarterly variables in Table 1. Accordingly, . Let denote the time mark of low-frequency data. , represents the j-th high-frequency observation of . For example, if , marks the third-month value of the high-frequency consumer confidence index in a specific quarter . After standardizing, we can stack the data into an mixed-frequency matrix . Assuming it follows a VAR(p) process, an MF-VAR(p) model can be established as

(1)

As a result, conducting a Granger causality test for confidence and the macroeconomy is equivalent to testing the parameters in Equation (1). For example, the null hypothesis that consumer confidence (GDP) is not a Granger cause of GDP (consumer confidence) cannot be rejected if and only if ().

The reduced form of Equation (1) is . To test the multiperiod relationships between the high-frequency and low-frequency variables, such as consumer confidence and GDP, the model can be expanded to the MF-VAR model, where refers to the number of periods:

(2)

where , , and . Testing the causal relationships between mixed-frequency variables over h periods is equivalent to testing whether the following constraints hold:

where is a selection matrix with full row rank and is the coefficient of the MF-VAR(p, h) model. is the constraint vector. The test statistic can be constructed by

(3)

where is the estimated ordinary least squares coefficients. is the consistent estimator of variance.

Under the null hypothesis, the p-value of the statistic can be calculated through Bootstrap by

(4)

where is the sample size. is the indicator function. At a given significance level , if , the null hypothesis is rejected, suggesting the existence of Granger causality among the variables.

In our case, after difference and detrending, we can stack series in a mixed-frequency matrix (omitting the time mark):

where , , and indicate the first, second and third monthly consumer confidence index in a quarter , respectively. The same for other variables. Based on this matrix, the MF-VAR model displayed in Equation (2) can be established. Then, the Wald statistic and corresponding p-value defined by Equations (3) and (4) can be calculated with Bootstrap to test the Granger causality among mixed-frequency variables. According to this statistic, we can test a series of null hypotheses, such as that consumer confidence is not the Granger cause of GDP, and GDP is not the Granger cause of consumer confidence.

3.2.2. The reverse unconstrained mixed data sampling model.

To utilize mixed-frequency data more effectively, Ghysels et al. [51] proposed the MIxed DAta Sampling (MIDAS) model based on distributed lag models. A linear combination of high-frequency variables explains a low-frequency variable in this model. This model requires a form of nonlinear least squares estimation, which increases the computational costs substantially in specifications with more than one high-frequency explanatory variable. Thus, it works well only in a univariate context where the frequency of the dependent variable is low and that of the independent variable is high, for example, quarterly to monthly. To avoid this limitation, Foroni et al. [49] introduced the reverse unconstrained mixed data sampling model (RU-MIDAS). In our case, the dependent variable is the high-frequency consumer confidence index, whereas the independent variables contain low-frequency macroeconomic variables such as GDP and urban unemployment rate. Therefore, RU-MIDA is more appropriate.

According to Foroni et al. [49], a RU-MIADS model can be represented as

(5)

where and are high- and low-frequency variables, respectively. is the potential high-frequency counterpart of . is the high-frequency lag operator, i.e., . is the ratio of sampling frequencies of and . . . is a high-frequency lag operator polynomial that characterizes the aggregation mechanism between low-frequency and . That is, . is a high-frequency lag operator polynomial satisfying the condition that contains only . is white noise.

In practice, though, the parameters of Equation (5) are unknown, meaning that cannot be determined precisely. Hence, empirically, we will use approximate RU-MIDA models based on linear lag polynomials, such as

(6)

where the orders of and , and , respectively, are large enough to make white noise.

The estimation of the RU-MIDAS is straightforward since it is a linear model. Hence, the parameters of the model in Equation (6) can be estimated by ordinary least squares for each value of , and the lag orders and can be selected using standard information criteria such as the Akaike information criterion and Bayesian information criterion. Moreover, the RU-MIDAS model in Equation (6) can be extended easily to allow additional high-frequency or low-frequency explanatory variables.

In our case, to decompose the confidence index, referring to Harrison and Weder [36] and Chen and Tang [13], we set the model as

(7)

where is the ratio of sampling frequencies. is the i-th month within a certain quarter. ) and are the regression coefficients of variables and their lagged terms, respectively. and represent the optimal lag orders for low-frequency and high-frequency variables, respectively. and are constant and random disturbances, respectively.

4.1. Descriptive statistics and correlation analysis

Table 2 presents the descriptive statistics of the variables. Combining Fig 2, we can find that the variables exhibit similar trends. This is preliminary evidence about the interaction between confidence and the macroeconomy. Specifically, due to the impacts of the global financial crisis and the COVID-19 pandemic, these variables significantly decreased in 2008 and 2020, though the consumer confidence index dropped sharply in April 2022.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Descriptive statistics.

https://doi.org/10.1371/journal.pone.0332909.t002

Before the formal analysis, we calculated pairwise Pearson and Spearman coefficients. The results (S2 Table) show that the purchasing managers’ index, the real estate climate index, interest rate, Shanghai stock composite index, and GDP (CPI, urban unemployment rate) are positively (negatively) correlated with the consumer confidence index. This is further evidence of the close relationship between confidence and the macroeconomy.

4.2. Granger causality tests

According to the AIC, the optimal lag order for the model displayed in Equation (2) is 2. The Wald statistic and corresponding p-values defined by Equations (3) and (4) are calculated with the Bootstrap method. The replication is set to 1000. The results are presented in Table 3.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Results of the mixed-frequency Granger causality test.

https://doi.org/10.1371/journal.pone.0332909.t003

It is found that, at the 10% significance level, the real estate climate index, GDP, and urban unemployment rate are Granger causes of consumer confidence, suggesting that changes in the real estate market, GDP, and urban unemployment rate will lead to changes in consumer confidence. Meanwhile, consumer confidence is a Granger cause of the CPI, indicating an intuitive conclusion that consumer confidence significantly influences consumer goods prices. Lastly, we cannot reject the null hypothesis that consumer confidence is not a Granger cause of GDP and urban unemployment rate in our sample. The reason may be a nonlinear Granger causality relationship between confidence and GDP and urban unemployment rate (Yuan et al. [32]).

4.3. The decomposition of confidence

Based on the AIC, the optimal lag orders in Equation (6) are determined. The estimation results of RU-MIDAS are presented in Table 4. As stated in Table 1, the consumer confidence index and real estate climate index are first-order differenced. The results show that the lagged changes in consumer confidence can explain itself significantly. Positive changes in the previous consumer confidence will lead to negative changes in the current period. In other words, the changes in consumer confidence proceed negatively despite the raw levels growing on the whole, despite the level shift effect of COVID-19 (see Fig 1). Second, CPI has a significant negative impact on changes in consumer confidence, indicating that as prices grow, the changes in consumer confidence will decline. This means that inflation is undesired among consumers. Third, changes in the real estate market positively impact changes in consumer confidence in the current and lagged periods, suggesting that the real estate market’s prosperity will significantly boost consumer confidence. This result is consistent with the basic facts – China’s macroeconomy’s rapid development has been driven considerably by the real estate market in recent years. Lastly, the purchasing managers’ index, interest rate, and Shanghai stock composite index are hardly essential factors affecting confidence. The reason is that these variables have progressed so steadily in recent years that consumers think they will not fluctuate significantly and become insensitive to these variables.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 4. Estimation results of RU-MIDAS.

https://doi.org/10.1371/journal.pone.0332909.t004

For low-frequency variables, GDP positively affects changes in consumer confidence, suggesting the evident and close nexus between consumer confidence and the macroeconomy, which is supported by most existing research. Lagged urban unemployment rate significantly negatively affects the change in consumer confidence. This is because employment and, thus, income correlate with confidence directly. These conclusions are mainly aligned with the mixed-frequency Granger causality test and most existing literature, such as Soric et al. [14].

4.4. The effect of animal spirits on the macroeconomy

The residual component that cannot be explained by macroeconomic fundamentals in the decomposition is the “animal spirits” (Fig 3). It fluctuates around zero and reaches a minimum in July 2022 due to the sluggish economic recovery following the COVID-19 pandemic. Additionally, it suffers from two fluctuation periods: one from early 2010 to the end of 2014, attributed to the spread of the financial crisis, and another from early 2017 to the end of 2022, influenced by Sino-US trade disputes and global COVID-19.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Animal spirits series.

The X-axis indicates the date (month), and the Y-axis indicates the level of animal spirits.

https://doi.org/10.1371/journal.pone.0332909.g003

We performed impulse response analysis to explore the relationship between “animal spirits” and the macroeconomy. Fig 4 shows the graph of the orthogonalized impulse response functions. It is found that most macroeconomic variables respond to the innovation from animal spirits in step 2 and beyond, indicating that the macroeconomy reacts to shocks from animal spirits slowly. And more importantly, most error bands include zero, meaning that most impulse responses are statistically insignificant. This means that shocks from animal spirits did not lead to noticeable fluctuations in China’s macroeconomy. This is because we have extracted the fundamental confidence from consumer confidence. The remaining is a purely irrational component. And whether under the rational expectation (Muth [21]), the bounded rationality (Simon [24]), or the imperfect rational expectations of information (Angeletos [25]), purely irrational expectations will not be the dominant factor in economic development solely.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Impulse response functions of the macroeconomic variables to “animal spirits”.

See Table 1 for definitions of variables. The horizontal axis tells the number of periods. The vertical axis indicates the values of the orthogonalized impulse response functions. The dotted lines indicate 95% confidence intervals.

https://doi.org/10.1371/journal.pone.0332909.g004

4.5. Robustness check