Ag Markets This Week Peak Trading Research

Ag Markets January 25, 2021

The macroeconomic environment is a neutral trading factor for agriculture futures today. Inflation expectations remain strong at an 8-year high.~However, energy markets are weak and the US dollar has strengthened against the Chinese yuan, ruble and Brazil (not good for agriculture products).

This week will be focused on Q4 corporate earnings, and the data calendar includes the FOMC and US GDP:

Wednesday: FOMC interest rate decision, Fed Chair Powell press conference
Thursday: US Q4 GDP (+4. 2% expected), jobless claims

Fund positioning

Fund positioning: Funds still hold large long positions across the agriculture sector after modest position adjustments over the weekend. The most "overvalued and overbought" agricultural products are spring wheat, Kansas wheat and corn.

Seasonality:

This week will be a big seasonal transition for agriculture futures. The #reflation trend of early January is coming to an end, with February and March seasonals firmly negative.

Most relevant questions this week

Funds trimmed their long positions last week, what catalysts will motivate funds to cover further?

1.) USD strength. The dollar is rising against key ag currencies such as BRL, RUB and CNY. Watch how the dollar moves this week, especially around the FOMC and US GDP. The dollar is key - charts for this week below.

2.) Seasonal traction. February and March seasonals turn sharply negative. Overbought markets look the most vulnerable, such as Kansas wheat, sugar No. 11 and Arabica coffee.

3.) Change in bullish fundamental scenario. Improving weather and Brazil's harvest countdown are negative factors. When will end users turn to buy? How will the spot market react to the firm bottom?

Final Chart of the Week A weaker dollar has been a big tailwind for agricultural prices over the past 9 months. There are two big USD catalysts this week: FOMC comments on Wednesday and US GDP data on Thursday.

Dave Whitcomb

Ag Markets January 18, 2021

Agricultural Price Drivers This Week

Hedge funds are long agricultural futures at record levels.
Fundamentals are tailwinds: weather in SA remains unsettled, China buying despite rising prices, Russian wheat export tax starting March 1st, and corn stocks tightening at last week's WASDE.
Macro clouds loom, including a stronger dollar and weaker energy prices.

Macro

The macro environment has lost some momentum, making this week a neutral trading day for agriculture futures. Inflation expectations remain strong (good for agriculture futures), while the US dollar has surged to a three-week high and energy markets have weakened (bad for agriculture futures). This is a big week for US politics and central banking, with President Biden's inauguration, possible violence in Washington D. C. and state capitals, President Biden's $1. 9 billion stimulus bill, new executive orders, and President Trump's impeachment (postponement?

Today: Martin Luther King Jr. Day, US markets closed

Tuesday: Yellen testifies before the Senate Banking Committee (in USD)
Wednesday: President Biden's inauguration
Brazil & Canada interest rate policy decisions~Thursday: ECB & BOJ policy decisions (focus on USD), US unemployment claims, Philly Fed index
Friday: US home sales
Fund Positioning

Fund positioning

Agricultural price seasonals this week are generally positive...

Seasonality:

What will trigger funds to liquidate record long positions?

1.) More dollar strength. The US dollar downtrend has been a big tailwind for agricultural futures over the past 9 months. If the US dollar strengthens again, it will drag agricultural futures down. This is a big macro week and macro-agricultural correlations are still hot. Keep an eye on the US dollar.

2.) Some change in bullish fundamental scenario. Now that the January WASDE is over, watch for signs of spot markets, Chinese buying and rationing. Too early for Brazilian harvest pressure.

3.) #Reflation is losing steam. Jobless claims rose and retail sales fell last week. Both suggest a slowing economic recovery and less growth-led inflationary pressure. The reflation trade is an early January phenomenon and inflows into commodity inflation hedges will likely slow in the coming days. Focus on gold, copper and Bitcoin.

When will the price seasonal turn bearish? Seasonal begins to change after this week. Seasonal in February is likely to be negative. Next month's seasonal will be a major threat to hedge fund positioning.

This week's chart this week will be a big week of Macro input, led by the inauguration ceremony of the President of Wednesday and a new presidential decree. The correlation between the macro input and the grain and oi l-seeding market is a strong plus, and this week, watch the US dollar, especially the BRL and CNY on CNY.

Note: This weekly "Agricultural Market of the Week" ends in February. If you would like this article in the future, INSIGHT@peakTradingResearch. com.

Hedge funds are recorded in agricultural futures for the January Wasde report on Tuesday.

These fund traders are convinced of lon g-term agricultural products by combining fundamentals (the strength of the actual market) and the price factors of no n-fundamentals (#reflation flow).

Dave Whitcomb

Ag Markets January 11, 2021

Wasde is a bearish data (eg, the rise in stock prices) and / or US dollars are the biggest threats to this week's larg e-scale long position.

Macro

The macro environment has reduced today's transaction input of today's agricultural futures due to the recent US dollar. The euro, gold and bitcoin are all compressed down.

Macro

This week, US inflation indicators will be announced:

Wednesday: Consumer Price Index (CPI) (Forecast: +0. 4 from the previous month

Thursday: Number of unemployment insurance applications, Powell FRB Chairman's lecture

Friday: US and Retail Sales Inflation Data
Fund positioning
In a Friday COT positioning report, the no n-commercial trader net long position was +1, 067, 077, a record high. The funds are a record long, a record long, and a record long throughout the grain market.

Fund positioning

Seasonal:

The price season in January is generally good and turns bearish in February.

Seasonality:

While the fund has recorded a record long, what is the opportunity to make the fund headed to liquidation?

When will the seasonality turn bearish? When will investors stop focusing on the reflation trade?

The seasonality will start to turn next week onwards. The February seasonality is firmly negative and fund selling has been going since early February 2011.

Could the macro environment prompt fund liquidation?

Yes, especially after the big WASDE report. Keep an eye on USD and CPI/PPI inflation data this week.

Charts of the Week Tuesday sees the release of the USDA Agriculture Data for January (WASDE). Hedge funds are preparing to go record long on agricultural futures, buoyed by the strength of the physical market and the January #reflation trend.

To try our industry-leading quantitative agriculture research, contact us at insight@peaktradingresearch. com.

Agriculture markets are receiving a big boost from a favorable macroeconomic environment, and investors will remain bullish on the reflation trade in the coming weeks.

Argentina’s suspension of new corn exports, a strong physical market, and continued weather and production uncertainty in South America are further supporting grain and oilseed markets. Calendar curve inverses are rising and hedge fund positions are at record levels entering this week.

Dave Whitcomb

Ag Markets January 4, 2021

Macro Environment

The macro environment is a strong tailwind for agriculture futures, driven by rising inflation expectations, a weak dollar, record stock market gains, and bullish sentiment in China. The correlation between macro indicators and agriculture futures is strong and will remain so through February.

Macro

This week is a big one macro-wise with Georgia elections on Tuesday and the NFP release on Friday:

Monday: OPEC+ production increase meeting

Tuesday: Georgia state senatorial runoff elections

Wednesday: FOMC minutes, ADP, Congress certifies Biden's election
Thursday: US unemployment claims
Friday US nonfarm payrolls (forecast: +50, 000, unemployment rate 6. 8)
Fund positioning
CFTC's COT report after the holidays will likely record hedge fund net long positions across agricultural commodities. Funds today are extending long positions in Kansas wheat, soybean meal, and corn.

Fund positioning

Price seasonals are positive in the first week of January as investors focus on the reflation trade and South American weather markets. Peak seasonal heatmap is a sea of green (chart below for this week).

Seasonality:

Non-fundamental: #January saw the reflation trade take center stage, with all commodities rising. Hedge funds are long at record levels, but with the positive macro environment and strong seasonality in January, now is not the time to be contrarian in fund positioning*. Keep an eye on USD (especially CNY and Brazil), oil (OPEC+ today) and equities.

Fundamentals: The end of the Argentine port strike should take some pressure off oilseeds, but new export restrictions on corn are boosting grains (following a similar move on wheat from Russia just three weeks ago). Forecast is dry for Argentina, moisture levels are still catching up in Brazil. SA weather market to continue until harvest gets going in February. WASDE report for January is released next Tuesday.

Chart of the Week #Reflation Trades Popular in January. Price seasonality in agriculture, energy and metals commodity markets reflects these flows. Peak's seasonal heatmap shows a sea of green in agriculture markets for the next few weeks.

If you would like to receive a trial of our industry-leading quantitative commodity research, please contact us at insight@peaktradingresearch. com.

What is happening in the economy right now? The future is uncertain, but so is the present. Policymakers operating in this environment need some way to understand what is happening now (i. e., the current economic situation). This need for timely information was most evident during the COVID-19 crisis, where the current situation changed rapidly and policymakers had to make decisions under great economic uncertainty. Policymakers are further hampered by the fact that gross domestic product (GDP), the most comprehensive measure of economic activity, is released with a significant delay. In fact, the full impact of the first major lockdown on economic activity, which occurred in the June 2020 quarter, was not realized until the release of national accounts data in early September 2020, more than two months after the base period.[1] [1] This lag limits the value of GDP to policymakers as a measure of the current state of the economy. Moreover, GDP is often revised in subsequent quarters, further limiting its usefulness to policymakers to assess the current situation.

However, they often do not have as comprehensive a scope or coverage as traditional indicators of economic activity such as GDP. And while these partial indicators help fill information gaps, the signals they provide are often noisy. Moreover, an indicator may be useful in one context but not another. For example, the unemployment rate is considered an important indicator of economic activity, but during the COVID-19 crisis, the Australian government introduced the “JobKeeper” program to keep workers employed, thereby curbing the rise in unemployment caused by the lockdown. [2] During this period, the underemployment rate was thought to give a more accurate signal. Given this, it is not clear which indicator policymakers should emphasize, and in the case of multiple indicators, how much importance each indicator should receive. The answer to such a decision is usually subjective and varies over time and between policymakers.

Research Discussion Paper – RDP 2024-04 Nowcasting Quarterly GDP Growth during the COVID-19 Crisis Using a Monthly Activity Indicator

What is needed is a method of combining the available part indicators in a systematic way, smoothing noise and clarifying fundamental signals. The most common tool that realizes this is the dynamic factor model (DFM). DFM is a dimension reduction technique that can summarize fluctuations common to the entire panel of time series data [3]. [3] In Australia, Gillitzer, Kearns and Richards (2005) conducted initial research to find the usefulness of factor models in monitoring economic activities. They have created two matching indicators that summarize quarterly data and monthly data. Both indicators were estimated using no n-parametric methods developed by Stock and Watson (2002) and Forni et al (2000). [4] Following this was Sheen, Trück and Wang (2015), which introduced a daily economy circulation index based on the research of aruoba et al (2009). Their method uses parametric estimation technology, which contains a stat e-o f-spatial model estimated using a kalman filter. What is needed is a method of combining the available part indicators in a systematic way, smoothing noise and clarifying the fundamental signal. The most common tool that realizes this is the dynamic factor model (DFM). DFM is a dimension reduction technique that can summarize fluctuations common to the entire panel of time series data [3]. [3] In Australia, Gillitzer, Kearns and Richards (2005) conducted initial research to find the usefulness of factor models in monitoring economic activities. They have created two matching indicators that summarize quarterly data and monthly data. Both indicators were estimated using no n-parametric methods developed by Stock and Watson (2002) and Forni et al (2000). [4] Following this was Sheen, Trück and Wang (2015), which introduced a daily economy circulation index based on the research of aruoba et al (2009). Their method uses parametric estimation technology, which contains a stat e-o f-spatial model estimated using a kalman filter. What is needed is a method of combining the available part indicators in a systematic way, smoothing noise and clarifying fundamental signals. The most common tool that realizes this is the dynamic factor model (DFM). DFM is a dimension reduction technique that can summarize fluctuations common to the entire panel of time series data [3]. [3] In Australia, Gillitzer, Kearns and Richards (2005) conducted initial research to find the usefulness of factor models in monitoring economic activities. They have created two matching indicators that summarize quarterly data and monthly data. Both indicators were estimated using no n-parametric methods developed by Stock and Watson (2002) and Forni et al (2000). [4] Following this was Sheen, Trück and Wang (2015), which introduced a daily economy circulation index based on the research of aruoba et al (2009). Their method uses parametric estimation technology, which contains a stat e-o f-spatial model estimated using a kalman filter.

Our research fills the literature gap between the use of a factor in monitoring and now casting in Australia. These two issues are related to each other, and it is wise to develop a framework that can achieve both purposes at the same time, as it is similarly important for policy proppons. At that time, by incorporating the recent developments of factor modeling and now casting, it is based on past research in Australia. First, develop Australia's monthly activity indicators (MAI). MAI aims to provide policy creators a more prominent snapshot in the current economic situation. We use the "true" DFM and the information obtained from the data set of 30 monthly forecast variables selected for the ability to explain the movement of the first quarterly GDP growth rate announced Is to achieve this by summarizing. [11] This is an important progress compared to pr e-research, as links to the variables that are interested in the estimation of DFM and improve the estimation and predictable ability of the factor (Baii and NG). 2008; Bulligan, Marcellino and Venditti 2015). We also extend the hard threshold preliminary step of the target predictive variable developed by Bai and NG (2008) when estimating the factor model to a mixed frequency setting. Furthermore, the method used to estimate MAI is longer than the competitors created by Gillitzer et al (2005) and Sheen et a (2005), which can be used to use unbalanced datasets. It means that you can consider a more widespread collections over time. Our research fills the gap in the literature that exists between the use of a factor in monitoring and now casting in Australia. These two issues are related to each other, and it is as important for policy creators, so it is wise to develop a framework that can achieve both purposes at the same time. At that time, by incorporating the recent developments of factor modeling and now casting, it is based on past research in Australia. First, develop Australia's monthly activity indicators (MAI). MAI aims to provide policy creators a more prominent snapshot in the current economic situation. We use the "true" DFM and the information obtained from the data set of 30 monthly forecast variables selected for the ability to explain the movement of the first quarterly GDP growth rate announced Is to achieve this by summarizing. [11] This is an important progress compared to pr e-research, as links to the variables that are interested in the estimation of DFM and improve the estimation and predictable ability of the factor (Baii and NG). 2008; Bulligan, Marcellino and Venditti 2015). We also extend the hard threshold preliminary step of the target predictive variable developed by Bai and NG (2008) when estimating the factor model to a mixed frequency setting. Furthermore, the method used to estimate MAI is longer than the competitors created by Gillitzer et al (2005) and Sheen et a (2005), which can be used to use unbalanced datasets. It means that you can consider a more widespread collections over time. Our research fills the literature gap between the use of a factor in monitoring and now casting in Australia. These two issues are related to each other, and it is wise to develop a framework that can achieve both purposes at the same time, as it is similarly important for policy proppons. At that time, by incorporating the recent developments of factor modeling and now casting, it is based on past research in Australia. First, develop Australia's monthly activity indicators (MAI). MAI aims to provide policy creators a more prominent snapshot in the current economic situation. We use the "true" DFM and the information obtained from the data set of 30 monthly forecast variables selected for the ability to explain the movement of the first quarterly GDP growth rate announced Is to achieve this by summarizing. [11] This is an important progress compared to pr e-research, as it shows that it links variables that are interested in the estimation of DFM and improves factors' estimation and predictive ability (Baii and NG). 2008; Bulligan, Marcellino and Venditti 2015). We also extend the hard threshold preliminary step of the target predictive variable developed by Bai and NG (2008) when estimating the factor model to a mixed frequency setting. Furthermore, the method used to estimate MAI is longer than the competitors created by Gillitzer et al (2005) and Sheen et a (2005), which can be used to use unbalanced datasets. It means that you can consider a more widespread collections over time.

Unlike previous nowcasting studies in Australia that focused only on quarterly frequency data (e. g., Australian Treasury 2018; Panagiotelis et al. 2019), we conduct the first investigation of nowcasting in Australia using a mixed-frequency modeling framework. We exploit the high-frequency information content of the MAI within a factor-augmented unrestricted MIDAS (MIxed Data Sampling) model (FA-U-MIDAS). [12] We evaluate the model’s ability to nowcast first-release quarterly GDP growth rates using a recursive out-of-sample evaluation exercise covering a 34-year period (1988:Q2-2022:Q2), longer than prior studies such as Gillitzer and Kearns (2007), Australian Treasury (2018), and Panagiotelis et al. (2019). Moreover, because we use monthly data, we are able to generate four nowcasts for each quarterly GDP growth observation as new monthly data become available across quarters. Finally, as in previous evaluations, we use a standard benchmark forecasting model with sample mean and AR(1) process for comparison.

The results show that incorporating monthly information reduces the estimated mean squared error and allows for more accurate forecasts compared to the benchmark model. The improvement over the benchmark model (sample mean and AR(1) model) is also found to be statistically significant. [13] Crucially, during the COVID-19 crisis, the forecast accuracy of the model using monthly data was greatest compared to the benchmark model that relied only on quarterly data, highlighting the benefit to policymakers of having timely information. Our results also support previous findings suggesting that the forecast accuracy of models varies with economic conditions (see Chauvet and Potter (2013), Siliverstovs (2020), Jardet and Meunier (2022)).

We first detail the methodology for constructing the MAI in Section 2. In Section 3, we discuss how we use the MAI to forecast quarterly GDP growth and the procedure for conducting out-of-sample estimations before concluding in Section 4. Some additional results are provided in the Appendix.

Apply DFM to the monthly dataset to build MAI. The dataset consists of a series showing a statistically significant relationship with the quarterly GDP growth rate. Before explaining the DFM method and estimated result, we will first explain the monthly dataset and how to select the series included in the dataset.

In relation, other methods for creating a dataset from the selected series have been proposed in order to generate more accurate predictions by returning to the factor. Unlike the methods proposed by Boivin and NG (2006), these methods use a dataset that is based only on the series that shows that the "target" variable has been predictable. We recommend that you estimate the factor model. The important thing is to explicitly consider the s o-called "target predictive variable" that is not considered by other methods. Two prominent strategies include "hard" and "software" threshold processing, which determines which variable to extract from which variable (Bai and NG 2008). In the hard threshold process, the predictive variable is ranked based on the pr e-test procedure, and those that do not meet a certain standard are discarded from the data set. In the soft threshold processing, a part of the hig h-ranked predictive variable is retained, and the order of the predictive variable depends on the used soft threshold processing rules. BAI and NG (2008) indicates that factors extracted from the targeted variable datasets may bring out excellent prediction performance. [15] Apply DFM to the monthly dataset to build Mai. The dataset consists of a series showing a statistically significant relationship with the quarterly GDP growth rate. Before explaining the DFM method and estimated result, we will first explain the monthly dataset and how to select the series included in the dataset.

2. Monitoring Activity Using a Combination of Targeted Monthly Indicators

In relation, other methods for creating a dataset from the selected series have been proposed in order to generate more accurate predictions by returning to the factor. Unlike the methods proposed by Boivin and NG (2006), these methods use a dataset that is based only on the series that shows that the "target" variable has been predictable. We recommend that you estimate the factor model. The important thing is to explicitly consider the s o-called "target predictive variable" that is not considered by other methods. Two prominent strategies include "hard" and "software" threshold processing, which determines which variable to extract from which variable (Bai and NG 2008). In the hard threshold process, the predictive variable is ranked based on the pr e-test procedure, and those that do not meet a certain standard are discarded from the data set. In the soft threshold processing, a part of the hig h-ranking predictive variable is maintained, and the order of the predictive variable depends on the used soft threshold processing rules. BAI and NG (2008) indicates that factors extracted from the targeted variable datasets may bring out excellent prediction performance. [15] Apply DFM to the monthly dataset to build MAI. The dataset consists of a series showing a statistically significant relationship with the quarterly GDP growth rate. Before explaining the DFM method and estimated result, we will first explain the monthly dataset and how to select the series included in the dataset.

2.1 Monthly activity dataset

In relation, other methods for creating a dataset from the selected series have been proposed in order to generate more accurate predictions by returning to the factor. Unlike the method proposed by Boivin and NG (2006), these methods use datasets that are only in the series that indicates that the "target" variables have predictive variables. We recommend that you estimate the factor model. The important thing is to explicitly consider the s o-called "target predictive variable" that is not considered by other methods. Two prominent strategies include "hard" and "software" threshold processing, which determines which variable to extract from which variable (Bai and NG 2008). In the hard threshold process, the predictive variable is ranked based on the pr e-test procedure, and those that do not meet a certain standard are discarded from the data set. In the soft threshold processing, a part of the hig h-ranked predictive variable is retained, and the order of the predictive variable depends on the used soft threshold processing rules. BAI and NG (2008) indicates that factors extracted from the targeted variable datasets may bring out excellent prediction performance. [15]

First, we created an “extended” dataset that includes 53 monthly partial indicators covering various aspects of the Australian economy. Following Bańbura and Rünstler (2007), we divided these series into three main categories: “hard” (30%; includes series covering key indicators of activity such as the labour market), “soft” (36%; includes survey indicators that tend to be more timely than the hard series), and “financial” (34%; includes series such as interest rates, equity prices, and commodity prices). When available, we include both aggregate and disaggregated measures in our dataset (i. e. total credit and its subcomponents). However, the method we used to estimate the DFM is robust to the inclusion of aggregate and individual series. [16] [16] The dataset covers the sample period 1978:M2 to 2022:M9 and is influenced by the number of series available early in the sample. [17] However, some series in the dataset are relatively new and therefore have late starts and early ends. Before using the dataset, we transform all series to be stationary and standardize them to have zero mean and unit variance, as is common in the factor modeling literature. Each series is stationary by taking logarithms and/or first differences as appropriate (see Table A1 for details). When standardizing, rather than using the full sample mean, we follow Kamber, Morley and Wong (2018) and perform “dynamic demeaning” of each series using rolling 20-year backward-looking estimates of the sample mean as a way to control for potential structural breaks in the central tendency of each series over the sample period covered by the dataset. The decision to use a 20-year window (rather than a 10-year window as in Kamber et al. (2018)) was based on the fact that the length of the Australian business cycle was arguably longer than other regions until the COVID-19-induced recession in 2020. [18]

Since our main goal is to create a monthly activity indicator to monitor the economy with a higher frequency than is currently possible and to forecast quarterly GDP growth in the near future, we follow Bai and Ng (2008) and implement a preselection strategy on our extended dataset to remove uninformative predictors related to quarterly GDP growth. Since our dataset is unbalanced, we use their hard threshold strategy. This involves running a series of separate regressions of the target on one predictor variable. Each regression includes a set of controls consisting of an intercept and a lag of the target variable that are the same in all regressions. Then, the predictors are ranked in descending order by the magnitude of the coefficient t-statistic for each predictor. Predictors with a test statistic below some specified threshold significance level are discarded. [19],[20]

The method we use to estimate the DFM is robust to model misspecification. Therefore, we can argue that there is no need to apply preselection to the dataset, since the model assigns appropriate weights to each series (see Bańbura et al (2013)). However, the factors extracted from the extended dataset are, by construction, linear combinations of all series in the dataset. Some of these series may not be very informative for quarterly GDP growth, but they will have some effect on the model output, even if it is small. That is, there are no series whose weightings are zero. Therefore, it makes sense to focus only on the subset of series that prove to be informative for quarterly GDP growth. [21]

Instead of using the current release version of the GDP, which is a combination of the first data, revision data, and complete revision data (Stone and Wardrop 2002), use the first version of GDP in accordance with the recommendation of Koenig and Dolmas and Piger (2003). Lee et al 2012). Extend the hard threshold algorithm of Bai and NG (2008) to the mixed frequency setting. This is because our predictive variables are monthly, while target variables are quarterly. In this situation, it is typical to execute a certain amount of tim e-based consolidation, such as the quarterly average (that is, the average of each quarter is the average of the three monthly objectivations in each quarter). However, this can lead to potential information loss. Instead, the monthly series is converted into a quarterly series by accumulating the first, second and third months of each quarter as three separate quarterly affiliates. [twenty two]

Because it has three prediction variables (that is, the first month, the second month, the third month), as well as one like Bai and NG (2008) and Bulligan et al (2015). The same T-statistical amount cannot be implemented to test the significance and rank the affiliate as in both sides. Instead, the WALD statistical amount calculated using the HAC Robust coordination matrix is used to verify the significance of all three series at once. As a control, it includes a set of seven indicators in the period of 2020: Q2 to 2021: Q2, 2021: Q4, because the cut and the sample cover the COVID-19 crisis period [23]. 23] The index variable was included in explaining the COVID-19 crisis so as not to affect the test results and affiliated rankings. [24] When executing each regression, the sample length of the subordinate variable is adjusted according to the sample length of each predictive variables depending on the affiliate. Our expansion data set is already relatively small in international standards, so we use a significant limit of 10 % to measure significance. [25] This is higher than the standard 5 %, but it helps to ensure a valid subset of an extended data set.

The result of the hard threshold process is a dataset of 30 variables from the original 53 variable extension dataset. [26] Among the three categories, 'Soft' is dominant in 13 series (43%), 'Financial' is 9 (30%), and 'Hard' is 8 series (27%). FIG. A1 shows the 30 series by category, and is ranked by the threshold critical value (dashed line) according to the wald statistics. The target number of predictive variables datasets is comparable to the minimum value proposed by Bai and NG (2008), slightly larger than the 24 series used in the empirical application of Bańbura et al (2013). It is slightly smaller than the 37 series used by Treasury (2018). Furthermore, Panagiotelis et al. (2019) stated that there is no advantage in taking into account a larger information set than 20 to 40 variables when predicting the Australian macro economy, such as the quarterly GDP growth rate. . [27]

DFM is a general statistical model for summarizing and predicting the common (linear) fluctuation contained in the panel of time series data. The important problem of these pr e-research is that it is not a true DFM as Bai and Wang (2015). Instead, we estimate MAI using the general format of DFM defined as follows:

2.2 Constructing the monthly activity indicator using a dynamic factor model

(1) y t = ∑ i = 0 s λ λ λ λ λ t, ε t ~ ε t ~ ε t ~ ε ∑ ∑ ∑ ∑ ∑ ∑ ∑ φ φ φ φ φ φ φ φ φ,,,,,,, ( , Q)

Here

t_{Is a weak regular target forecast variable n × 1 vector, f}t_{Is a weak regular target forecast variable n × 1 vector, f}T-I_{I = 0, 1. S and t = 1. T. By combining this factor to the load, a common fluctuation scale shared between the series in the dataset is obtained. The dynamic factor is modeled as a VAR (P) process, which is the Q x-Q matrix of φ i (all roots are outside the unit of unit). The number of dynamic factors is Q (F)}t_{Is a weak regular target forecast variable n × 1 vector, f}).

t_{Is a weak regular target forecast variable n × 1 vector, f}t_{Is a weak regular target forecast variable n × 1 vector, f}We estimate DFM by the sem i-lover (QMLE). [30] The estimation is performed by the expected value (EM) algorithm and consists of two parts. First, the kalman filter and Rauch-tung-striebel (RTS) smoother recursion to estimate the data given factor ("e-step"). Next, the model parameters are calculated by using the estimated factor in the previous step to maximize the ant i-compatible likelihood of regression ("M step"). For this, it is necessary to r e-cast the formula (1) to the following state spatial expression:

(2) Y t = λ f T + ε t f-t = φ f T-1 + g η t

The measurement equation takes the form of a static factor model (Stock and Watson 2002) with the static factor of R = Q (S + 1). If k = max (p, s + 1), f

t_{Is a weak regular target forecast variable n × 1 vector, f}Before estimating DFM, it is necessary to first identify the four important features. i) Number of dynamic factors (Q), II number of dynamic load (S), III) Factors VAR's lag number (p), IV) It is the "named factor" required for identification. 。 . Second, it is assumed that the dynamics of the factors are captured by the var (p) process. [28] BAI and WANG (2015) is the first dynamics to make this specification a true dynamics model because the biggest difference between dynamic and static factor analysis is the maximum difference between dynamic factor analysis and static factor analysis. He claims to be a source. [29]

2.2.1 Determining the optimal DFM specification

We estimate DFM by the sem i-lover (QMLE). [30] The estimation is performed by the expected value (EM) algorithm and consists of two parts. First, the kalman filter and Rauch-tung-striebel (RTS) smoother recursion to estimate the data given factor ("e-step"). Next, the model parameters are calculated by using the estimated factor in the previous step to maximize the ant i-compatible likelihood of regression ("M step"). For this, it is necessary to r e-cast the formula (1) to the following state spatial expression:

(2) Y t = λ f T + ε t f-t = φ f T-1 + g η t

The measurement equation takes the form of a static factor model (Stock and Watson 2002) with the static factor of R = Q (S + 1). If k = max (p, s + 1), f

Is a dynamic factor and its rug QK x 1 vector, λ is an N × QK matrix with dynamic factor load, φ is a QK × QK companion matrix, and G is a QK × q selector matrix. The advantage of using a state spatial modeling framework is that it can easily and efficiently respond to unbalanced datasets. For more information about the parameters and estimation procedures we use, see Hartigan and Wright (2023).

Before estimating DFM, it is necessary to first identify the four important features. i) Number of dynamic factors (Q), II number of dynamic load (S), III) Factors VAR's lag number (p), IV) It is the "named factor" required for identification. 。 Second, it is assumed that the dynamics of the factors are captured by the var (p) process. [28] BAI and WANG (2015) is the first dynamics to make this specification a true dynamics model because the biggest difference between dynamic and static factor analysis is the maximum difference between dynamic factor analysis and static factor analysis. He claims to be a source. [29]

2.2.2 Estimation results

(2) Y t = λ f T + ε t f-t = φ f T-1 + g η t

The measurement equation takes the form of a static factor model (Stock and Watson 2002) with the static factor of R = Q (S + 1). If k = max (p, s + 1), f

We use information standards developed by Hallin and Liška (2007) to determine the number of dynamic factors. This suggests that there is only one common dynamic factor in the targeted variable data set (see Fig. A4). [31] To set the number of dynamic factor load, we follow the strategy implemented in Luciani (2020). This uses the fact that a dynamic factor model with a Q individual can be r e-cast as a static factor model with R = Q (s + 1) as described above. In fact, a subset with a balance of the target predictive variable data set is taken, and the percentage of fluctuations described from the first R individual value from the coordinated distribution queue at the same time, and the average in the frequency grid. Compare the percentage of fluctuations described from the dynamic value of the first Q individual from the spectrum density matrix (see Forni et al (2000) and Brillinger (1981) for details). The purpose here is to find a place where there is a closely match between these two scale. Verifying Table 1 that one dynamic value (ie, q = 1) explains the same amount of fluctuations as the three static specific values (that is, R = 3), and thus suggest S≈2. In addition, Luciani (2020) argues that in the case of S = 2, the targeted predictable datasets can be loaded into a dynamic factor in a thre e-month window. This is interesting because this window is equivalent to the first quarter.

Table 1: Explained variations_{Is a weak regular target forecast variable n × 1 vector, f}By comparing the amount of fluctuations described from the expansion dataset and the targeted dataset set, it is possible to "sharpen the factor structure" by reducing the sample size. You can check if the proposal (2006) is correct. Focusing only on the first dynamic value is about 52 % (not shown) in the amount of fluctuation described in a wel l-balanced subset of the expansion dataset, and is explained in a screening dataset in advance. Lower than percentage of fluctuations (Table 1). Therefore, by deleting a series that is not considered to be useful in explaining the movement of the quarterly GDP growth rate, the signal noise ratio of common dynamic factors increased.

Since there is only one common factor, the dynamics of the factors obey the AR process, not the VAR process. In order to determine the number of lags in the AR process, set it to 1 (ie, p = 1) based on AIC._{Is a weak regular target forecast variable n × 1 vector, f}Figure 1 shows the (optimal) filter estimates of the MAI for the sample period: from 1978:M2 to 2022:M9. [33] The MAI reveals three periods of relatively weak activity corresponding to past recessions in Australia, most recently since the COVID-19 crisis (i. e., 1982, 1989-1991, and 2020). Indeed, the decline in the MAI level during this period is the largest ever observed in this series. However, its duration is very short compared to the other two periods, and is concentrated in June 2020.

Figure 1: Monthly Activity Indicators

Note: The shaded bars are the start dates of the Australian Sarm rule recession, defined as an increase of 3/4 percentage points in the 3-month moving average of the unemployment rate from its lowest value in the previous 12 months._{Is a weak regular target forecast variable n × 1 vector, f}In order to understand the motion of MAI over time, it is necessary to quantify individual affiliates. This is something that can not be directly done with no n-parametric methods such as PCA. These contributions are not provided as part of the estimation procedure, but can be obtained from the stat e-o f-th e-art expression of the formula (2). First, take out the equation of the condition of the model and rewrite this formula from the viewpoint of the update equation from the Kalman filter:_{Is a weak regular target forecast variable n × 1 vector, f}Here, k

Is a kalman gain in time T and the dimension of R x N. Equation (3) means that the estimation of the common factor in time T is the prediction step (based on the information in the T-1) and the linear combination of the update step based on the prediction error weighted by the kalmangain. The second part is a contribution to each affiliate at each point (see Sheen et al (2015)).

Next, D

Shows the RX N matrix that contributes to the affiliated series, and can use the formula (3) to obtain the formula of the update steps in each series: [35]

(4) d t = k t ⊙ [(y T-λ (φ F T-1))) '↪ sm_2297 ι R].

Equation (4) focuses on the formula (4) by focusing on the f-t = φ F T-1 + d t ι (here, the ι m is the row vector with an element specified by M). It is related to 3).

The first line is T = 1. T is a common dynamic factor F

3. Predicting Quarterly GDP Growth Using the MAI

Indicates a unique contribution to each line. Figure 2 shows a data category (hard, soft, finance) of the contribution to the Kalman filter estimated value from January 2000 to July 2022.

3.1 Modelling mixed frequency data

Figure 2: MAI contribution by data category

Note: Contribution to condition renewal in Kalman filter recursion.

Figure 2 shows that soft data has greatly contributed to MAI updates. Interesting in this observation is that GFC was generally considered a financial crisis. However, according to the breakdown of MAI by data category in Fig. 2, the weakness of the MAI that occurred between GFCs in Australia is mainly due to the decrease in software data, mainly in centments. It can be found that it is a basis. [36] Financia l-based affiliates have contributed very much at this time.

During the GFC, the economic environment was very uncertain, and both consumers and companies have expressed many pessimism. However, it turned out that the Australian government's very larg e-scale financial response, a very aggressive monetary easing by the RBA, and the rapid increase in product demand from China, and the serious concerns of the serious recession were early. 。 Until early 2018, steady rise in MAI was the main cause of soft data. More recently, the dramatic movement observed in the MAI during the COVID-19 crisis has contributed from both categories of hard data and soft data, and the financial data category is a relatively small contribution. represents the R x N matrix that contributes to the affiliated affiliated, and can use the formula (3) to obtain the expression of the update steps in each series: [35]

(4) d t = k t ⊙ [(y T-λ (φ F T-1))) '↪ sm_2297 ι R].

Equation (4) focuses on the formula (4) by focusing on the f-t = φ F T-1 + d t ι (here, the ι m is the row vector with an element specified by M). It is related to 3).

The first line is T = 1. T is a common dynamic factor F

Indicates a unique contribution to each line. Figure 2 shows a data category (hard, soft, finance) of MAI's contribution to the Kalman filter estimated value from January 2000 to July 2022.

Figure 2: MAI contribution by data category

Note: Contribution to condition renewal in Kalman filter recursion.

During the GFC, the economic environment was very uncertain, and both consumers and companies have expressed many pessimism. However, it turned out that the Australian government's very larg e-scale financial response, a very aggressive monetary easing by the RBA, and the rapid increase in product demand from China, and that the severe recession was too early. 。 Until early 2018, steady rise in MAI was the main cause of soft data. More recently, the dramatic movement observed in the MAI during the COVID-19 crisis has contributed from both categories of hard data and soft data, and the financial data category is a relatively small contribution. Shows the RX N matrix that contributes to the affiliated series, and can use the formula (3) to obtain the formula of the update steps in each series: [35]

(4) d t = k t ⊙ [(y T-λ (φ F T-1))) '↪ sm_2297 ι R].

Equation (4) focuses on the formula (4) by focusing on the f-t = φ F T-1 + d t ι (here, the ι m is the row vector with an element specified by M). It is related to 3).

t< 2 , 3 , 4 , 5 >The first line is T = 1. T is a common dynamic factor F

Indicates a unique contribution to each line. Figure 2 shows a data category (hard, soft, finance) of the contribution to the Kalman filter estimated value from January 2000 to July 2022.

Figure 2: MAI contribution by data category	Note: Contribution to condition renewal in Kalman filter recursion.					Figure 2 shows that soft data has greatly contributed to MAI updates. Interesting in this observation is that GFC was generally considered a financial crisis. However, according to the breakdown of MAI by data category in Fig. 2, the weakness of the MAI that occurred between GFCs in Australia is mainly due to the decrease in software data, mainly in centments. It can be found that it is a basis. [36] Financia l-based affiliates have contributed very much at this time.		During the GFC, the economic environment was very opaque, and both consumers and companies have expressed many pessimism. However, it turned out that the Australian government's very larg e-scale financial response, a very aggressive monetary easing by the RBA, and the rapid increase in product demand from China, and the serious concerns of the serious recession were early. 。 Until the beginning of 2018, steady rise observed in MAI was the main cause of soft data. More recently, the dramatic movements observed in the MAI during the COVID-19 crisis were contributed from both hard data and soft data categories, and the financial data category was only a relatively small contribution.
	This analysis reveals potential issues that users of the MAI as an indicator of activity should consider. While soft data such as surveys have the advantage of being very timely compared to hard data categories, they can also provide spurious signals (Aylmer and Gill 2003). Furthermore, Roberts and Simon (2001) conclude that the information content provided by survey data such as sentiment indicators is, at best, a rough summary of the economic situation. However, they note that in some cases, a linear combination of survey indicators (as in the DFM) is a good compromise.		As mentioned earlier, the DFM used to construct the MAI has been shown to be robust to misspecification, including conditional heteroscedasticity and “fat tails” (i. e., outliers), when factors are extracted from many variables (see Doz et al. (2012) and Bańbura et al. (2013)). However, as Figure 1 makes clear, the COVID-19 crisis had substantial, previously unobserved effects on many of the series in our target predictor dataset. Moreover, Maroz, Stock and Watson (2021) document that the COVID-19 crisis caused a significant temporary change in previously observed patterns of co-movement across a panel of monthly time series data for the United States. Although they use a different model than ours, it is still important to check the robustness of our model estimates.		As mentioned earlier, the DFM used to construct the MAI has been shown to be robust to misspecification, including conditional heteroscedasticity and “fat tails” (i. e., outliers), when factors are extracted from many variables (see Doz et al. (2012) and Bańbura et al. (2013)). However, as Figure 1 makes clear, the COVID-19 crisis had substantial, previously unobserved effects on many of the series in our target predictor dataset. Moreover, Maroz, Stock and Watson (2021) document that the COVID-19 crisis caused a significant temporary change in previously observed patterns of co-movement across a panel of monthly time series data for the United States. Although they use a different model than ours, it is still important to check the robustness of our model estimates.
	Compared to the discovery of MAROZ et al. (2021), one of the reasons for the small effects observed in our case show the extreme movement as their datasets during the COVID-19 crisis. Maybe because it wasn't. In fact, according to their reports, there are affiliates that have fallen by more than 275 standard deviations. In our dataset, the maximum decline was much smaller (see A6). Furthermore, considering that the decline observed in MAI (PC) during the COVID-19 period was relatively small, COVID in the next section, when looking at the quarterly GDP growth rate, COVID throughout the line. It is appropriate to claim that it is necessary to include the COVID-19 period in order to correctly estimate the impact o f-19.	Figure 3: Impact on the MAI estimate of the COVID-19 crisis	Compared to the discovery of MAROZ et al. (2021), one of the reasons for the small effects observed in our case show the extreme movement as their datasets during the COVID-19 crisis. Maybe because it wasn't. In fact, according to their reports, there are affiliates that have fallen by more than 275 standard deviations. In our dataset, the maximum decline was much smaller (see A6). Furthermore, considering that the decline observed in MAI (PC) during the COVID-19 period was relatively small, COVID in the next section, when looking at the quarterly GDP growth rate, COVID throughout the line. It is appropriate to claim that it is necessary to include the COVID-19 period in order to correctly estimate the impact o f-19.	Figure 3: Impact on the MAI estimate of the COVID-19 crisis	Compared to the discovery of MAROZ et al. (2021), one of the reasons for the small effects observed in our case show the extreme movement as their datasets during the COVID-19 crisis. Maybe because it wasn't. In fact, according to their reports, there are affiliates that have fallen by more than 275 standard deviations. In our dataset, the maximum decline was much smaller (see A6). Furthermore, considering that the decline observed in MAI (PC) during the COVID-19 period was relatively small, COVID in the next section, when looking at the quarterly GDP growth rate, COVID throughout the line. It is appropriate to claim that it is necessary to include the COVID-19 period in order to correctly estimate the impact o f-19.	Figure 3: Impact on the MAI estimate of the COVID-19 crisis	Compared to the discovery of MAROZ et al. (2021), one of the reasons for the small effects observed in our case show the extreme movement as their datasets during the COVID-19 crisis. Maybe because it wasn't. In fact, according to their reports, there are affiliates that have fallen by more than 275 standard deviations. In our dataset, the maximum decline was much smaller (see A6). Furthermore, considering that the decline observed in MAI (PC) during the COVID-19 period was relatively small, COVID in the next section, when looking at the quarterly GDP growth rate, COVID throughout the line. It is appropriate to claim that it is necessary to include the COVID-19 period in order to correctly estimate the impact o f-19.	Figure 3: Impact on the MAI estimate of the COVID-19 crisis
0:2	(5) y t = β 0 + β 1 W ( L 1 / m ; θ ) x t ( m ) + ∈ t	Where W ( L 1 / m ; θ ) = Σ k - 0 K - 1 W ( k ; θ ) L k / m and L 1/m are high-frequency lag operators such that L 1 / m x t ( m ) = x t - 1 / m ( m ), where m denotes the high sampling frequency of the explanatory variables (e. g., m = 3 when x is monthly and y is quarterly). The intercept is designated β 0 , while the coefficient β 1 captures the overall effect of the high-frequency variable x on y and can be identified by normalizing the function W ( L 1 / m ; θ ) to sum to 1. We assume that the residuals ∈ t are zero-mean, constant-variance iid sequences. Finally, K is the maximum lag length of the high-frequency regressors included.	The way in which the MIDAS model achieves a concise representation is through the lag coefficients of W ( k ; θ ), which represents a set of weights as a function of a small-dimensional vector of j parameters θ = ( θ 0 , θ 1 , . . , θ j ), j ≪ K . A commonly used function for heuristic applications is the normalized exponential Almon lag function of Ghysels et al (2004):	(6) W ( k : θ ) = θ 0 exp ( θ 1 k + θ 2 k 2 + . + θ j k j ) Σ k = 1 K exp ( θ 1 k + θ 2 k 2 + . + θ j k j )	And the normalization beta function of GhySels et al (2007) is given as follows:	(6) W ( k : θ ) = θ 0 exp ( θ 1 k + θ 2 k 2 + . + θ j k j ) Σ k = 1 K exp ( θ 1 k + θ 2 k 2 + . + θ j k j )	Note: FIG. B1 shows an example of a doubl e-term weighting function for different parameters. Since both weighted functions are highly no n-linear, the MIDAS model must be estimated by the no n-linear minimum square method. The alternative specification proposed by Foroni et al (2015) is 'UNRESTRICTED MIDAS' (U-Midas). This method can be estimated by OLS by remaining the hig h-frequency lag coefficient in no n-restraint. [43] FORONI et al (2015) is often more desirable than U-Midas (that is, restricted) Midas (R-Midas) when modeling the quarter and monthly data because the M is small. It shows that. This reflects the fact that if the number of lags to be modeled is relatively small, the complexity of more parameters must be reduced. [44] A U-Midas model with one explanatory variable is given:	na
0:3	(8) Y t = β 0 + b (L 1 / m) x t (m) + ∈ t	(6) W ( k : θ ) = θ 0 exp ( θ 1 k + θ 2 k 2 + . + θ j k j ) Σ k = 1 K exp ( θ 1 k + θ 2 k 2 + . + θ j k j )	Before moving to the MIDAS model specifications for the nau cast of the quarterly GDP growth rate using MAI, you need to determine the specifications to be used. One way to deal with these two issues is to select the optimal model from the viewpoint of parameters constraints and lags based on the i n-sample model conformity using the information standard.	(6) W ( k : θ ) = θ 0 exp ( θ 1 k + θ 2 k 2 + . + θ j k j ) Σ k = 1 K exp ( θ 1 k + θ 2 k 2 + . + θ j k j )	Before moving to the MIDAS model specifications for the nau cast of the quarterly GDP growth rate using MAI, you need to determine the specifications to be used. One way to deal with these two issues is to select the optimal model from the viewpoint of parameters constraints and lags based on the i n-sample model conformity using the information standard.	(6) W ( k : θ ) = θ 0 exp ( θ 1 k + θ 2 k 2 + . + θ j k j ) Σ k = 1 K exp ( θ 1 k + θ 2 k 2 + . + θ j k j )	). [50] The results are shown in Table 2. The BIC strongly prefers the U-MIDAS specification with maximum lag K = 5, closely followed by the U-MIDAS model with maximum lag K = 6. Indeed, all U-MIDAS models outperform the two R-MIDAS models except when the maximum lag is two (K = 2). In this case, the R-MIDAS model with a normalized exponential Almon polynomial weighting function with two parameters (j = 2) is preferred.	na
0:4	In addition to the comparison of BIC-based models, it is also possible to verify the empirical validity of the polynomed weighted function used in the R-MIDAS specification under the standard assumption by the Wald type. The returnless hypothesis is that function restriction is effective. Therefore, rejection of the returnless hypothesis means that function restrictions are not supported by data. In this measurement standard, only one R-MIDAS model specification matches the data, which corresponds to the normalized index almon multiple weight function of K = 2 and J = 2.	(6) W ( k : θ ) = θ 0 exp ( θ 1 k + θ 2 k 2 + . + θ j k j ) Σ k = 1 K exp ( θ 1 k + θ 2 k 2 + . + θ j k j )	Before moving to the MIDAS model specifications for the nau cast of the quarterly GDP growth rate using MAI, you need to determine the specifications to be used. One way to deal with these two issues is to select the optimal model from the viewpoint of parameters constraints and lags based on the i n-sample model conformity using the information standard.	(6) W ( k : θ ) = θ 0 exp ( θ 1 k + θ 2 k 2 + . + θ j k j ) Σ k = 1 K exp ( θ 1 k + θ 2 k 2 + . + θ j k j )	Normalization beta	(6) W ( k : θ ) = θ 0 exp ( θ 1 k + θ 2 k 2 + . + θ j k j ) Σ k = 1 K exp ( θ 1 k + θ 2 k 2 + . + θ j k j )	j = 2	na
0:5	j = 3	(6) W ( k : θ ) = θ 0 exp ( θ 1 k + θ 2 k 2 + . + θ j k j ) Σ k = 1 K exp ( θ 1 k + θ 2 k 2 + . + θ j k j )	Before moving to the MIDAS model specifications for the nau cast of the quarterly GDP growth rate using MAI, you need to determine the specifications to be used. One way to deal with these two issues is to select the optimal model from the viewpoint of parameters constraints and lags based on the i n-sample model conformity using the information standard.	(6) W ( k : θ ) = θ 0 exp ( θ 1 k + θ 2 k 2 + . + θ j k j ) Σ k = 1 K exp ( θ 1 k + θ 2 k 2 + . + θ j k j )	Bic	(6) W ( k : θ ) = θ 0 exp ( θ 1 k + θ 2 k 2 + . + θ j k j ) Σ k = 1 K exp ( θ 1 k + θ 2 k 2 + . + θ j k j )	Bic	na
P-value

3.2 Out-of-sample prediction comparison

Bic

P-value

486. 88_t0. 86_t ₊₁492. 05

0. 00

505. 09

0. 00

491. 35_t500. 63_t0. 00< 3 , 2 , 1 , 10 >492. 06

0. 00

492. 06

0. 00

437. 71

501. 95

0. 00

492. 06

0. 00

492. 08

0. 00

417. 46

502. 19

0. 00

492. 06

0. 00	492. 24	FC	M1	M2	M3	QA
0. 00
422. 58
Note The variable J is the number of parameters of the multiple weight weight function used in MIDAS regression. The P-value is for testing a hypothesis of whether or not the constraint of the MIDAS coefficient implicated by the multilema weighted function is supported by data. The bold value indicates the best model.	This section evaluates the performance of the Midas model Naucast, which incorporates the MAI, by comparing a pseud o-o f-the sample (OOS) comparison compared to the standard benchmark model. Based on the results of the model evaluation shown in Table 2, only the FA-U-MIDAS specification for navigating the quarterly GDP growth rate is examined. However, instead of K = 5 as suggested by BIC, K = 6 is decided. This selection is motivated by past research (see Koening et al (2003) and Leboeuf and Morel (2014)) and the last month (that is, the end of the real GDP growth rate, which is likely to affect the quarterly GDP growth rate. This is because it covers the thre e-month data that covers the quarter of the value of the values and the data for the first quarter of the nau casting). [51]	Like other studies, lon g-term predictions of quarterly GDP growth are not taken into account. This is because the prediction of the production growth rate for a longer period is known to be less reliable (for example, for the FA-Midas model, Marcellino and Schumacher (2010), and for the factor model, Bańbura et al. (2013), more common systematic evaluation is CHAUVET AND POTTER (2013)). Therefore, the method developed here should only be considered as a shor t-term prediction device.	The important advantage of Return to Midas is that the predictions can be made within the period, compared to other methods used to handle mixed frequency data (that is, time consolidation). Furthermore, continuous forecasts will incorporate MAI's new estimated values as the number of data available in the quarter increases. Assuming the timing of the data release as follows for simple. y	The quarterly quarterly GDP growth rate, y	Shows the next quarterly quarterly GDP growth rate. The first release of the T-quarter GDP contains data up to the T-1 quarter. MAI will be updated four times before the T +1 quarter of the GDP growth rate of the T period becomes available. The first estimation of MAI incorporating monthly data to T-1 is T-2/3 (that is, the first month of the quarter), the second MAI incorporated monthly data from T-2/3. The estimation is T-1/3 (that is, the second month of the quarter), and the third estimate of MAI incorporating monthly data from T-1/3 is T (that is, the end of the current quarter). Finally, the fourth estimation of MAI incorporating monthly data to T of T +1/3 (that is, the first month of the next quarter). With the timing of the monthly update of these MAIs, you can create four forecasts in the quarterly GDP growth rate in the current quarter. These are forecast (FC), II) Nau cast (M1), III) in the first month (III) Nau cast (M2), IV). 。 [52] See Fig. 4 for visual summaries.	Figure 4: GDP NowCast timeline	Based on this, the general FA-U-Midas model used in OOS evaluation is as follows:
(9) Y t = β 0 + ∑ k = I k-1 β K K L K / M X T (m) + ∈ T	Here, y	Is the GDP growth rate in the first quarter, x	Is Mai, i	Changes according to the quarterly monthly data flow (that is, four predictions in the order of FC, M1, M2, and M3). Therefore, as the new monthly estimated value of MAI is created during the quarter, the specifications of the FA-U-Midas model change as the number of regression variable increases. For example, in the case of I = 3 (that is, FC), the FA-U-Midas model for the current quarterly GDP growth is composed of three-month data on cuts and the previous quarter of MAI:	(10) Y t = β 0 + β 1 x T-1 + β 2 x T-4 /3 + β 3 x T-5 /3 + ∈ T	Or, when I = 0, the model is expanded to include additional data for MAI from the current term (reflecting a full model):	(11) Y t = β 0 + β 0 + β 1 x T + β 2 x T-1 /3 + β 3 + β 4 x X-1 + β 5 X T-4 /3 + β 6 x T-5 /3 + ∈ t
In order to make the notation easier, both formulas have removed an upper character (M) from the X variable. During the OOS evaluation period, FA-U-Midas model specifications should be fixed to this general form. [53] Furthermore, compare the four FA-U-MIDAS model specifications with the two standard models used in the past OOS prediction/ NowCasting evaluation. These include specimen average and AR (1) processes (see Australian Treasury (2018) and Panagiotelis et Al (2019) for the average specimen, and Gillitzer and Kearns (2007) for AR (1) for AR (1) processes). The specimen average is shown to be a formidable prediction model of the quarterly GDP growth rate (Panagiotelis et al 2019) and functions as a benchmark model in our comparison. In addition, we will consider a model based on the quarterly (QA) measurement of MIA as a cross check. [54] The QA model is similar to the M3, including the time tallying of the MAI in the current quarter and the different rug value in the previous quarter.	In order to evaluate the performance of various models, all samples are divided into estimated su b-samples and evaluation sub samples to perform recursive estimates, predictions, and nowCasting. The estimated sample covers 1978: Q2-1988: Q1 period (that is, as Panagiotelis et al (2019)), expands one quarter of each year at a time, and is expanded for one quarter at a time, and each model parameter each time. Will be estimated again. The evaluation sample is between 1988: Q2 and 2022: Q2 (that is, p = 137). For each quarter of the evaluation sample, I would like to calculate the prediction and three naucasts according to the monthly information set. For example, for the initial evaluation quarter 1988: Q2, the prediction is calculated using data up to 1988: Q1 (FC), and 1988: m4 (m1), 1988: m5 (m2), 1988: Q2 (m3). I want to calculate NauCast. At this point, the sample average and AR (1) prediction are also calculated. Fig. 5 shows the predicted values of each model in the evaluation sample.	Figure 5: Fort y-quarter real GDP growth prediction	(11) Y t = β 0 + β 0 + β 1 x T + β 2 x T-1 /3 + β 3 + β 4 x X-1 + β 5 X T-4 /3 + β 6 x T-5 /3 + ∈ t	Source Abs; Author Calculation; Lee et al (2012).	Considering that the quarterly GDP growth rate is not persistent, the sample average and the prediction of AR (1) are very similar, but the AR (1) model has fallen COVID-19 in the second quarter of 2020. Although there was a quarterly lag, it was some excellent to predict. Since both models were estimated using quarterly data, it was not possible to completely predict the significant decline in the second quarter of 2020 and the third quarter of 2020. In contrast, the models incorporating monthly information showed much better. The prediction of each model is similar in the period before the COVID-19 crisis, but all models show a remarkable difference. For example, the model M1 predicted the reduction of the quarterly GDP growth rate in the second quarter of 2020, but still had an error of about 2 %. On the other hand, model M3 has relatively low accuracy. This is surprising because the model M3 has added data from the second quarter of 2020 for two months, and has shown that having more timely data improves normal predictive accuracy. It should be something [55]. [55] One explanation is that the model M3 has two parameters that are estimated compared to the model M1, and the increase in estimation may have an effect on the accuracy of the model.	The success rate of the FC model and the QA model was high for the prediction of the major rise in the quarterly GDP growth rate in the third quarter of 2020. This is surprising because both models contain less information about the quarter than three "M" models. However, both models have less parameters to estimate compared to other models (three FCs and two QA), so they could be more accurate and improved the accuracy of both models.	Source Abs; Author Calculation; Lee et al (2012).
(12) R S ME = 1 P ∑ T = 1 P (Y T-Y ^ T) 2 < Span> Given that the quarterly GDP growth is not persistent, the sample average and AR (1) prediction are very good. Similar, but the AR (1) model was a little better for predicting the decline of COVID-19 in the second quarter of 2020, despite the quarterly lag. Since both models were estimated using quarterly data, it was not possible to completely predict the significant decline in the second quarter of 2020 and the third quarter of 2020. In contrast, the models incorporating monthly information showed much better. The prediction of each model is similar in the period before the COVID-19 crisis, but all models show a remarkable difference. For example, the model M1 predicted the reduction of the quarterly GDP growth rate in the second quarter of 2020, but still had an error of about 2 %. On the other hand, model M3 has relatively low accuracy. This is surprising because the model M3 has added data from the second quarter of 2020 for two months, and has shown that having more timely data improves normal predictive accuracy. It should be something [55]. [55] One explanation is that the model M3 has two parameters that are estimated compared to the model M1, and the increase in estimation may have an effect on the accuracy of the model.
Note The variable J is the number of parameters of the multiple weight weight function used in MIDAS regression. The P-value is for testing a hypothesis of whether or not the constraint of the MIDAS coefficient implicated by the multilema weighted function is supported by data. The bold value indicates the best model.	na	Using the standard tw o-square average square root error (RMSE) defined as follows, evaluate the point prediction of each model that was examined / index accuracy:	(12) R S ME = 1 P ∑ T = 1 P (Y T-Y ^ T) 2 Considering that the quarterly GDP growth rate is not durable, the prediction of the sample average and AR (1) are very similar. However, the AR (1) model was a little better in predicting the decline of COVID-19 in the second quarter of 2020, despite the quarterly lag. Since both models were estimated using quarterly data, it was not possible to completely predict the significant decline in the second quarter of 2020 and the third quarter of 2020. In contrast, the models incorporating monthly information showed much better. The prediction of each model is similar in the period before the COVID-19 crisis, but all models show a remarkable difference. For example, the model M1 predicted the reduction of the quarterly GDP growth rate in the second quarter of 2020, but still had an error of about 2 %. On the other hand, model M3 has relatively low accuracy. This is surprising because the model M3 has added data from the second quarter of 2020 for two months, and has shown that having more timely data improves normal predictive accuracy. It should be something [55]. [55] One explanation is that the model M3 has two parameters that are estimated compared to the model M1, and the increase in estimation may have an effect on the accuracy of the model.	The success rate of the FC model and the QA model was high for the prediction of the major rise in the quarterly GDP growth rate in the third quarter of 2020. This is surprising because both models contain less information about the quarter than three "M" models. However, both models have less parameters to estimate compared to other models (three FCs and two QA), so they could be more accurate and improved the accuracy of both models.	Using the standard tw o-square average square root error (RMSE) defined as follows, evaluate the point prediction of each model that was examined / index accuracy:	(12) R S ME = 1 P ∑ T = 1 P (Y T-Y ^ T) 2	During the whole sample period, the model M1 exceeds all other models in each of the three Horizon. Relatively, the model M1 RMSE is more than half of the average model model in the last three and ten years, and is less than thre e-quarters of the benchmark model in all samples. The QA model is the only model that has achieved the same performance in all sample Horizon. As described above in the relationship with Fig. 5, this result is mainly based on how well the Model M1 has successfully predicted a significant decline in the quarterly GDP growth that occurred in the second quarter of 2020. This can be supported by comparing the model RMSE in the previous period of COVID-19. Here, there is no model that surpasses others over all periods, as in the case of the COVID-19 period. In addition, all models of RMSE are remarkably low and closer. Models that incorporate monthly information are not so dominant. In fact, the three 'm' models are better than the average sample model, both in a short thre e-year Horizon or a lon g-time al l-sample Horizon. In contrast, the QA model suggests that all three Horizen exceed the benchmark model, and that the prediction using timely information has some benefits. However, it also suggests that there may be trad e-off between the size and accuracy of the model, especially in the relatively "normal" period.
(9) Y t = β 0 + ∑ k = I k-1 β K K L K / M X T (m) + ∈ T	na	Using the standard tw o-square average square root error (RMSE) defined as follows, evaluate the point prediction of each model that was examined / index accuracy:	AR (1)	Complete sample	D i-square average square root error	For the past three years	2. 72
In order to make the notation easier, both formulas have removed an upper character (M) from the X variable. During the OOS evaluation period, FA-U-Midas model specifications should be fixed to this general form. [53] Furthermore, compare the four FA-U-MIDAS model specifications with the two standard models used in the past OOS prediction/ NowCasting evaluation. These include specimen average and AR (1) processes (see Australian Treasury (2018) and Panagiotelis et Al (2019) for the average specimen, and Gillitzer and Kearns (2007) for AR (1) for AR (1) processes). The specimen average is shown to be a formidable prediction model of the quarterly GDP growth rate (Panagiotelis et al 2019) and functions as a benchmark model in our comparison. In addition, we will consider a model based on the quarterly (QA) measurement of MIA as a cross check. [54] The QA model is similar to the M3, including the time tallying of the MAI in the current quarter and the different rug value in the previous quarter.	na	2. 11	1. 24	1. 72	AR (1)	1. 24	1. 72
1. 52
422. 58
Note The variable J is the number of parameters of the multiple weight weight function used in MIDAS regression. The P-value is for testing a hypothesis of whether or not the constraint of the MIDAS coefficient implicated by the multilema weighted function is supported by data. The bold value indicates the best model.	0. 74	0. 74	1. 26	0. 87	all	0. 98	1. 01
(9) Y t = β 0 + ∑ k = I k-1 β K K L K / M X T (m) + ∈ T	0. 70	0. 78	0. 88	0. 70	0. 88	0. 70	0. 88
In order to make the notation easier, both formulas have removed an upper character (M) from the X variable. During the OOS evaluation period, FA-U-Midas model specifications should be fixed to this general form. [53] Furthermore, compare the four FA-U-MIDAS model specifications with the two standard models used in the past OOS prediction/ NowCasting evaluation. These include specimen average and AR (1) processes (see Australian Treasury (2018) and Panagiotelis et Al (2019) for the average specimen, and Gillitzer and Kearns (2007) for AR (1) for AR (1) processes). The specimen average is shown to be a formidable prediction model of the quarterly GDP growth rate (Panagiotelis et al 2019) and functions as a benchmark model in our comparison. In addition, we will consider a model based on the quarterly (QA) measurement of MIA as a cross check. [54] The QA model is similar to the M3, including the time tallying of the MAI in the current quarter and the different rug value in the previous quarter.	0. 45	0. 45	0. 82	0. 55	For the past 10 years	1. 04	0. 79
(12) R S ME = 1 P ∑ T = 1 P (Y T-Y ^ T) 2 < Span> Given that the quarterly GDP growth is not persistent, the sample average and AR (1) prediction are very good. Similar, but the AR (1) model was a little better for predicting the decline of COVID-19 in the second quarter of 2020, despite the quarterly lag. Since both models were estimated using quarterly data, it was not possible to completely predict the significant decline in the second quarter of 2020 and the third quarter of 2020. In contrast, the models incorporating monthly information showed much better. The prediction of each model is similar in the period before the COVID-19 crisis, but all models show a remarkable difference. For example, the model M1 predicted the reduction of the quarterly GDP growth rate in the second quarter of 2020, but still had an error of about 2 %. On the other hand, model M3 has relatively low accuracy. This is surprising because the model M3 has added data from the second quarter of 2020 for two months, and has shown that having more timely data improves normal predictive accuracy. It should be something [55]. [55] One explanation is that the model M3 has two parameters that are estimated compared to the model M1, and the increase in estimation may have an effect on the accuracy of the model.
Note The variable J is the number of parameters of the multiple weight weight function used in MIDAS regression. The P-value is for testing a hypothesis of whether or not the constraint of the MIDAS coefficient implicated by the multilema weighted function is supported by data. The bold value indicates the best model.	na	(10) Y t = β 0 + β 1 x T-1 + β 2 x T-4 /3 + β 3 x T-5 /3 + ∈ T	0. 57	all	1. 03	0. 89	0. 71
(9) Y t = β 0 + ∑ k = I k-1 β K K L K / M X T (m) + ∈ T	na	2. 11	In order to evaluate the performance of various models, all samples are divided into estimated su b-samples and evaluation sub samples to perform recursive estimates, predictions, and nowCasting. The estimated sample covers 1978: Q2-1988: Q1 period (that is, as Panagiotelis et al (2019)), expands one quarter of each year at a time, and is expanded for one quarter at a time, and each model parameter each time. Will be estimated again. The evaluation sample is between 1988: Q2 and 2022: Q2 (that is, p = 137). For each quarter of the evaluation sample, I would like to calculate the prediction and three naucasts according to the monthly information set. For example, for the initial evaluation quarter 1988: Q2, the prediction is calculated using data up to 1988: Q1 (FC), and 1988: m4 (m1), 1988: m5 (m2), 1988: Q2 (m3). I want to calculate NauCast. At this point, the sample average and AR (1) prediction are also calculated. Fig. 5 shows the predicted values of each model in the evaluation sample.	COVID-19 or earlier samples	D i-square average square root error	Figure 5: Fort y-quarter real GDP growth prediction	D i-square average square root error
In order to make the notation easier, both formulas have removed an upper character (M) from the X variable. During the OOS evaluation period, FA-U-Midas model specifications should be fixed to this general form. [53] Furthermore, compare the four FA-U-MIDAS model specifications with the two standard models used in the past OOS prediction/ NowCasting evaluation. These include specimen average and AR (1) processes (see Australian Treasury (2018) and Panagiotelis et Al (2019) for the average specimen, and Gillitzer and Kearns (2007) for AR (1) for AR (1) processes). The specimen average is shown to be a formidable prediction model of the quarterly GDP growth rate (Panagiotelis et al 2019) and functions as a benchmark model in our comparison. In addition, we will consider a model based on the quarterly (QA) measurement of MIA as a cross check. [54] The QA model is similar to the M3, including the time tallying of the MAI in the current quarter and the different rug value in the previous quarter.	na	(10) Y t = β 0 + β 1 x T-1 + β 2 x T-4 /3 + β 3 x T-5 /3 + ∈ T	0. 34	0. 33	0. 36	2. 11	For the past 10 years
0. 46

0. 47

3.3 Evaluating model performance during the COVID-19 crisis

0. 45

0. 43

0. 45

0. 46

0. 45

3.4 Assessing the predictive content of the MAI

all

0. 59_i0. 59< AR ( 1 ) FC,M1,M2,M3,QA >0. 64

0. 62

0. 61

0. 60

0. 56

4. Conclusion

Relative tw o-square average square root error

For the past three years

1. 00

1. 18

Appendix A: Additional Monthly Activity Indicator Details

A.1 Additional MAI dataset details

1. 15

Beyond the peak: How has the inflation crisis shaped the UK economy?

1. 10

1. 21

0. 98

For the past 10 years

1. 03

0. 98

UK hit hard compared to peers

0. 94

0. 97

1. 01

0. 97
all

1. 00

1. 09

1. 06

1. 05

Lower-income households struggled the most

1. 03

0. 96

1. 01

0. 97
As shown in Table 3, Horizon, which covers the COVID-19 crisis, had a significant divergence on the predictive accuracy of the model incorporating monthly information and quarterly information. Most of the results are due to a certain point: in June 2020, the first lockdown of COVID-19, which was obliged by the government, and the congestion of the economic activity that caused it. It is. Figure 6 shows the results of comparing the predictive errors of each model with the first published value of the actual quarterly GDP growth rate of the period. If the number is 1 or more, it means that the prediction error was larger than the actual GDP result, and if the numbers were less than 1, the prediction error was smaller than the actual GDP result.

Chart 6: Comparison of relative forecasting and predictive errors-The second quarter of 2020

Note: The error is a relative error to the first quarter GDP growth rate (-7 %) in the second quarter of 2020. Quarter GDP growth rate When comparing the results of the prediction / expectation related to the forecast, it is appropriate to focus only on the sample results before COVID-19. From this point of view, our results show a clear advantage of using higher frequency (monthly) data to predict data with lower frequency (quarter). The Australian Treasury (2018) and Panagiotelis et al (2019), which focus on quarterly data, do not consistently exceed the sample benchmark model. However, the Australian Finance model can exceed the specimen average model if all data in the quarter (the timing may be equivalent to our M3 and QA models). In contrast, all FA-U-MIDAS models except M3 have outperformed sample models on average over the past 10 years, and the QA version is full sample for the past three years (1988: Q2-2019: Q4) shows the outperation.

As shown in Table 3, Horizon, which covers the COVID-19 crisis, had a significant divergence on the predictive accuracy of the model incorporating monthly information and quarterly information. Most of the results are due to a certain point: in June 2020, the first lockdown of COVID-19, which was obliged by the government, and the conversion of the economic activity that resulted in the quarterly. It is. Figure 6 shows the results of comparing the predictive errors of each model with the first published value of the actual quarterly GDP growth rate of the period. If the number is 1 or more, it means that the prediction error was larger than the actual GDP result, and if the numbers were less than 1, the prediction error was smaller than the actual GDP result.

1. 21

Brits saved more than US households

Note: The error is a relative error to the first quarter GDP growth rate (-7 %) in the second quarter of 2020. When comparing the results of the prediction of the quarterly GDP growth rate with our results, it is appropriate to focus only on the sample results before COVID-19. From this point of view, our results show a clear advantage of using higher frequency (monthly) data to predict data with lower frequency (quarter). The Australian Treasury (2018) and Panagiotelis et al (2019), which focus on quarterly data, do not consistently exceed the sample benchmark model. However, the Australian Finance model can exceed the specimen average model if all data in the quarter (the timing may be equivalent to our M3 and QA models). In contrast, all FA-U-MIDAS models except M3 have outperformed sample models on average over the past 10 years, and the QA version is full sample for the past three years (1988: Q2-2019: Q4) shows the outperation.

As shown in Table 3, Horizon, which covers the COVID-19 crisis, had a significant divergence on the predictive accuracy of the model incorporating monthly information and quarterly information. Most of the results are due to a certain point: in June 2020, the first lockdown of COVID-19, which was obliged by the government, and the congestion of the economic activity that caused it. It is. Figure 6 shows the results of comparing the predictive errors of each model with the first published value of the actual quarterly GDP growth rate of the period. If the number is 1 or more, it means that the prediction error was larger than the actual GDP result, and if the numbers were less than 1, the prediction error was smaller than the actual GDP result.

Chart 6: Comparison of relative forecasting and predictive errors-The second quarter of 2020

Note: The error is a relative error to the first quarter GDP growth rate (-7 %) in the second quarter of 2020.

1. 21

Government debt not saved by inflation

Similar to the forecast results, the errors of models M2 and M3, which include more timely information, are also significantly larger than model M1. The performance of the QA model using a time-aggregated version of the MAI (i. e., 3-month average) appears to strike a compromise between the three MIDAS models. However, our main result that incorporating timely information can improve the forecast accuracy of the model during recessions is consistent with previous studies such as Clements and Galvão (2009) (the 2001 US recession), Schorfheide and Song (2015) (the impact of the 2008 GFC on US economic activity), and Jardet and Meunier (2022) (the impact of the COVID-19 pandemic on global GDP growth).

The relative RMSE results in the previous section show that models incorporating monthly information produce more accurate forecasts (smaller errors) compared to the baseline sample average model. However, crucially, it is important to compare the performance of models using a formal statistical test of equal forecast accuracy.

Because we are evaluating nested models (all models include an intercept), we cannot use the well-known Diebold-Mariano-West (DMW) t-type test of equal predictive accuracy. Instead, we follow the approach of Clark and McCracken (2005) and Clements and Galvão (2009) and implement a bootstrapped version of the MSE- F test of equality of mean squared errors (MSE) developed by McCracken (2007). [57] MSE

denotes the MSE from model i against

1. 21 1. 01

Here P is the number of predictions to be compared. A negative MSE-F means that the model I is not accurate compared to the specimen average model, while the positive MSE-F means that the model I is more accurate. The boot strap is used to calculate the P-value of the MS E-F test and progresses as follows. The average sample model is estimated using the first quarterly GDP growth rate (as clements and galvão (2009) recommended). From the model fit, the estimated fragments and residue distribution are dispersed, and multiple time series trajectory is simulated from the specimen average model that assumes Gaussian properties. [58] For each of the simulated time-series trajectory, apply the same recursive estimate and prediction steps used in actual data to calculate the MSE-F statistics of the reproduction. Mai is fixed in each replication. Set the total number of repetitions in the boot strap procedure to 1, 000. Experiented P-value is calculated as the ratio of MSE-F statistics from a simulation that is larger than the MSE-F statistical amount calculated using actual data. As we did in the RMSE comparison in Table 3, a boot-strap MSE-F test is performed for all samples, including the COVID-19 period, and the shorter sub-samples except for the COVID-19 period. The results are shown in Table 4.
Note The benchmark model is a specimen average. Experiented P-value is calculated by boot strap using 1, 000 repetition. Completely specimens are 1988: Q2-2022: Q2, COVID-19 or more specimens are 1988: Q2-2019: Q4. Bold is rejected the return of the hypothesis. And the test statistics of the substitute model specifications that are considered as the benchmark specification average model (that is, μ) are implemented using the following test statistics:

(13) M S E-F = P × (M S E μ-M S E i) M S E I

Here P is the number of predictions to be compared. A negative MSE-F means that the model I is not accurate compared to the specimen average model, while the positive MSE-F means that the model I is more accurate. The boot strap is used to calculate the P-value of the MS E-F test and progresses as follows. The average sample model is estimated using the first quarterly GDP growth rate (as clements and galvão (2009) recommended). From the model fit, the estimated fragments and residue distribution are dispersed, and multiple time series trajectory is simulated from the specimen average model that assumes Gaussian properties. [58] For each of the simulated time-series trajectory, apply the same recursive estimate and prediction steps used in actual data to calculate the MSE-F statistics of the reproduction. Mai is fixed in each replication. Set the total number of repetitions in the boot strap procedure to 1, 000. Experiented P-value is calculated as the ratio of MSE-F statistics from a simulation that is larger than the MSE-F statistical amount calculated using actual data. As we did in the RMSE comparison in Table 3, a boot-strap MSE-F test is performed for all samples, including the COVID-19 period, and the shorter sub-samples except for the COVID-19 period. The results are shown in Table 4.

Note The benchmark model is a specimen average. Experiented P-value is calculated by boot strap using 1, 000 repetition. Completely specimens are 1988: Q2-2022: Q2, COVID-19 or more specimens are 1988: Q2-2019: Q4. Bold is rejected the return of the hypothesis. Then, the test that is equal to the alternative model specifications that are considered as the benchmark specimen average model (that is, μ) is conducted using the following test statistics:

(13) M S E-F = P × (M S E μ-M S E i) M S E I

1. 21

Overall, these results mirror those of Chauvet and Potter (2013) and Siliverstovs (2020) regarding the change in forecast accuracy of quarterly GDP growth in the United States between economic expansions and recessions. In our case, the statistical evidence of the advantage of models incorporating more timely information over simple models based on quarterly information is largely due to the significant outperformance over the three-year period covering the COVID-19 crisis. In contrast, in more “normal” times, model forecasts incorporating monthly information fail to significantly improve the forecasts of the benchmark sample average model. This is not surprising given Figure A3. Australia’s quarterly GDP growth is serially uncorrelated. However, model QA, which includes current quarter information (albeit averaged), was able to consistently outperform the benchmark model in both sample periods. This suggests that there may be a trade-off between incorporating more timely information and increasing model complexity. One curious aspect of our results is that, unlike many other studies where GDP growth forecasts become more accurate as more quarterly data are available, our results show the opposite: they become less accurate. We speculate that this is related to the increased parameter uncertainty caused by estimating progressively larger models. Furthermore, the COVID-19 crisis may induce very large outliers (also known as "leverage points" because of their effect on the estimated regression fit), which can have a significant impact on parameter estimates.

Nevertheless, our results are encouraging news for policymakers. They show that by using a MIDAS-based model incorporating the timely MAI, it is possible to more accurately forecast Australia's quarterly GDP growth during times of crisis (such as the COVID-19 crisis), when accuracy is most needed. This is because the higher frequency information contained in the MAI allows the MIDAS model to detect sudden changes sooner, giving policymakers time to react.

Elim Poon - Journalist, Creative Writer

Last modified: 27.08.2024

Peak Trading Research assesses the latest COT report, examining alterations in price and open interest to provide an appraisal of net fund positioning daily. Technical analysis caveat; Summary. Introduction. Agriculture commodity markets are analyzed either technically or fundamentally. Fundamental analysis studies. The most 'expensive & overbought' markets across the ag complex are now spring wheat, kansas wheat, and corn.

Play

accepted