When data don't go your way, just change the rules, move the limits, or, best of all, introduce brand new data that will validate your assumptions. This has been demonstrated very well in Fukushima over the past month. Now it is coming out from Goldman's economic team, which is finding GDP to not be quite as amenable to "presenting" for client indoctrination purposes, due to its recent plunge from expectations (especially those of young master Hatzius, discussed here). As a result the Hatzius et al team have decided to launch an experiment in scrapping GDP as the key indicator of economic growth (or lack thereof) for those periods in which it is dropping, and instead will focus on the CAI, or the Current Activity Indicator: a synthetic Made In Goldman bogus indicator, which ignores the weak data, and emphasizes the good stuff. Brilliant. Goldman's recent addition to its economic team Zach Pandl explains. Elsewhere, Zero Hedge is launching a contest for the best abuse of the CAI acronym to explain what it really means...
CAI: A Measure for Tracking US Growth
GDP is the most common summary measure of activity, but it has several flaws. Most importantly, it excludes many important indicators (e.g. industrial production, ISM, payroll employment) and is not very timely. Alternative summary measures may be more valuable from a market perspective.
With this in mind, we have created a Current Activity Indicator (CAI), and find that it complements our other tools. In statistical jargon, the measure is defined as the “first principal component” of 24 real activity indicators, expressed in GDP-equivalent terms.
At present, the current activity measure carries a clear message: US growth likely had considerable momentum in late Q1. According to the CAI approach, the positive signal from the business surveys and from a few of the other “hard” indicators (e.g. hours worked) dominates negative news from expenditure-related data—consistent with our forecast that GDP growth will accelerate again in the second quarter.
Gross Domestic Product (GDP) is the most widely used summary measure of economic activity. GDP underpins governments’ budgets and provides the structure for most macroeconometric models. Data watchers spend considerable effort “tracking” its progress through time. But GDP has some serious flaws: it is only available quarterly and released at least a month after activity took place; initial estimates are based on incomplete data and therefore heavily revised; it is based primarily on expenditure figures even though other series, such as income-related data, may provide more reliable estimates; and “noise” in a few components, like trade and inventories, can have a large affect on short-term trends.
Given these drawbacks, other summary measures of current activity may be useful as well, particularly from a market perspective. With this in mind, we have created a US Current Activity Indicator (CAI), and find that it complements our existing tools.
We construct the CAI in three steps:
1. Model high-frequency indicators. Summary measures of economic activity are usually only available with a long lag. Because they incorporate data from many different sources, they are only as timely as the slowest indicator. For example, a proper measure of monthly GDP can only be calculated 6-7 weeks after the end of the month when the Census Bureau releases data on retail inventories. However, while we sometimes lack actual values for certain indicators, we always have expectations for those variables based on current information. Before the release of nonfarm payrolls, weekly jobless claims reports provide a rough sense of labor market conditions. Daily gas price data inform our views on the CPI, and chain-store results our expectation for retail sales. Our assessment of current conditions becomes less uncertain with the release of additional data, but even before that we are not completely in the dark.
We exploit this simple idea for our CAI. As a first step, we model the interrelationships between 24 weekly and monthly real activity indicators for the US economy. We then use this system to forecast each variable for periods when actual values are not yet available. The table below lists the indicators we include. There are three types: surveys, measures of real expenditures (which are used to calculate GDP), and other “hard” indicators. A few others (e.g. the Empire State index and gasoline prices) are included in the forecasting system but not in the final CAI.
2. Calculate common component. Variation in real activity indicators can be thought of as the sum of two parts: 1) a component common to all measures, which likely relates to the business cycle; and 2) an idiosyncratic component particular to each indicator. For example, both the ISM index and real disposable income vary over the business cycle, but only the latter is directly affected by changes in the tax code (at least over the short run). To capture this common trend, we calculate the first principal component of the 24 indicators. The first principal component is the time series that minimizes the proportion of the variation in the sample that is explained by idiosyncratic factors. The resulting principal component is a weighted average of the indicators expressed as normalized scores (i.e. with zero mean and unit standard deviation). Before complete data is available for all indicators, the principal component reflects a blend of actual values and model-based expectations.
3. Express in GDP-equivalent terms. GDP may have flaws, but it is still the way in which most people think about growth. In contrast, the units of our first principal component are not immediately intuitive. Therefore, as a final step, we express the first principal component in GDP-equivalent terms. The goal is not to forecast GDP – for which we think our “bean-count” modeling still works better – but to present the series in recognizable units. We first average the principal component into quarterly values and then regress it on real GDP growth and a constant (note that we use fully revised GDP growth, not originally reported/unrevised figures). We then apply the estimated coefficients to the monthly estimate of the first principal component. This final measure is our Current Activity Indicator: the first principal component of 24 indicators, expressed in GDP terms. The exhibit below shows the CAI along with quarterly GDP growth.
There is already a large literature on alternative summary measures of US activity. Our series is most akin to the Chicago Fed National Activity Indicator (CFNAI), originally developed by James Stock and Mark Watson (“Forecasting Inflation”, Journal of Monetary Economics, October 1999). The CFNAI is the first principal component of 85 real activity indicators related to production and output, the labor market, real expenditures and income, and business conditions. The simple correlation between the CFNAI and our measure is 92%. Differences include: 1) the specific indicators in the sample; 2) the fact that we express the series in GDP-equivalent terms; and 3) that we can calculate our indicator at any time. The CFNAI is published monthly around four weeks after the end of the reference month.
Other measures more loosely related to our CAI include the US Treasury’s Real-Time Forecasting System (RTFS) and the Arouba-Diebold-Scotti (ADS) Business Conditions Index published by the Philadelphia Fed (for RTFS, see “Real-Time Forecasting in Practice”, Business Economics, October 2003; for ADS, see “Real-Time Measure of Business Conditions”, Journal of Business and Economic Statistics, October 2009).
The CAI has a few notable statistical properties. First, it is dominated by surveys: the six component indicators with the largest weight are all surveys. Expenditure data and other “hard” indicators matter, but they are also influenced by more idiosyncratic factors – weather, worker strikes, tax law changes, seasonal adjustment distortions, etc. The typical “hard” indicator has about half the weight of the best of the surveys. In practice, this means that a one standard deviation change in a highly weighted survey indicator (e.g. the ISM) is worth about a two standard deviation change in a hard indicator (e.g. imports or consumer spending). Second, the CAI is a moderately better predictor of future GDP growth than current GDP, at least in-sample (at a one-quarter horizon, the regression standard error is 10% smaller). Academic research, such as the Stock and Watson paper cited above, has generally found favorable properties in principal component-based indicators (see also Ben Bernanke and Jean Boivin, “Monetary Policy in a Data-Rich Environment”, Journal of Monetary Economics, April 2003). Third, the CAI can be revised, both because its components are revised and, for the latest month, because we replace forecasted values with actual results as they are released.
At present, the current activity measure carries a clear message: US growth likely had considerable momentum in late Q1. Although our “bean-count” model of Q1 GDP suggests growth of 2.5% or lower, the CAI showed growth of 3.6% in February. For March the index is currently tracking growth of more than 4%, because available surveys remain strong and declines in other indicators (e.g. imports and housing starts) look likely to reverse. According to the CAI approach, the positive signal from the business surveys and from a few of the other “hard” indicators (e.g. hours worked) dominates negative news from expenditure-related data—consistent with our forecast that GDP growth will accelerate again in the second quarter.