Financial markets pay great attention to reported changes in key economic statistics, particularly when they are unexpected. For quantitative analysis, we introduce the concept of information state changes and the methods of aggregating them across time and indicators. We apply these to a few popular U.S. indicators and investigate how information state changes have affected the bond market. In line with theory, monthly changes in economic growth, inflation, and employment growth have all been negatively correlated with concurrent Treasury returns over the past 25 years. However, there has been subsequent payback: the correlation reverses for subsequent monthly Treasury returns. This supports the hypothesis that high publicity volatile indicators are easily “overtraded.” Cognitive biases may systematically exaggerate positioning toward the latest “surprises” or publicized changes.
The post below is based on Macrosynergy’s proprietary research.
Please quote as “de la Porte Simonsen, Lasse, Sueppel, Ralph and Tyagi, Palash, ‘Reported economic changes and the Treasury market: impact and payback,’ Macrosynergy research post, July 2024.”
A Jupyter notebook allows for the audit and replication of the research results. The notebook’s operation requires access to J.P. Morgan DataQuery to download data from JPMaQS, a premium service of quantamental indicators. J.P. Morgan offers free trials for institutional clients.
Also, an academic research support program sponsors data sets for relevant projects.
What are macro information states?
Macroeconomic information states are the latest published instances of macroeconomic metrics, such as GDP growth rates, short-term inflation trends, or unemployment rates. Their timestamps are real-time dates, different from the observation periods for which the economic events have been measured. For example, the information state of an inflation rate may be timestamped on the 6th of March, but refer to the observation period in February. Simply put, an information state is a measure of public knowledge, not of the actual occurrence of an economic event.
In the absence of insider trading, information states represent the macroeconomic information set of the market that underlies end-of-day prices. Due to this synchrony, information states are also valid indicators for assessing the predictive power of macroeconomic indicators with respect to subsequent market returns and for running realistic backtests. Macroeconomic information states that are particularly relevant for liquid financial asset prices are called macro-quantamental indicators. They are available from specialized data services, such as the J.P. Morgan Macrosynergy Quantamental System (“JPMaQS”). JPMaQS provides daily end-of-day (New York time) information states of major macroeconomic indicators that are commonly observed by financial investors for up to 40 countries, based on concurrent data “vintages.”
Previous Macrosynergy posts have provided evidence that financial markets are far from information efficiency with respect to economic trends and states, in accordance with the theory of rational inattention. By contrast, traders and analysts do pay great attention to reported changes in key economic indicators, particularly surprises. This particular focus has implications for the usage of such data as trading signals:
- Markets swiftly evaluate and possibly trade on information state changes of important indicators, such as those related to growth and inflation in the United States. Indeed, economic release calendars of services such as Bloomberg are designed for this aspect of economic data watching and make analysis very easy and user-friendly. This means that, unlike in the case of economic trends, there is no reason to assume gradual permeation of information state changes or surprises.
- Information states could still provide trading value through skilful aggregation of a range of indicators across time and different countries, including those with low publicity. As shown below aggregation is possible if one respects the basic structure and logic of time series of information state changes.
- Anecdotal and empirical evidence of this post suggests that markets may “overtrade” economic information state changes and surprises, depending on their publicity and the stability of the underlying indicators. This means that changes in popular but volatile indicators affect prices concurrently in the “right” direction, according to standard theory, but entail payback in subsequent periods.
What should we know about changes in information states?
Principally, changes in information states are merely the first differences in recorded time series of information states. However, to calculate these changes and use them correctly in trading strategies, one must consider a few important features:
- Changes in information states of economic indicators have multiple causes. Mostly, changes occur when the data set adds a new observation period or when statistics are being revised. Sometimes, changes also arise from updates of model parameters, such as nowcasting coefficients or seasonal adjustment factors. Finally, sometimes information states change because of changes in convention, such as the central bank’s preferred definition of core inflation.
- The frequency of information state changes depends on indicator calculation. Typically, the number of information state changes per year depends negatively on the length of the observation period, positively on the frequency of revisions of days other than release days, and positively on the number of sub-indicators used for calculation. For example, a quarterly bank lending survey score that is only revised on days of new releases would only give four information state changes per year. By contrast, the nowcast of a monthly GDP growth rate based on many high-frequency reports could give 150-200 changes per year. Generally, low-frequency updates have more individual impact than high-frequency updates.
- Information state changes are not all surprises. The concept of information state changes is a necessary step towards estimating economic surprises, but not a sufficient one. Surprises are information state changes that are not anticipated by the market. In practice, there are two factors that allow systematically predicting part of information state changes: early indicators and mathematical-statistical biases. Early indicators are neglected data that are released for an observation period prior to the main indicator that is followed by the market. Mathematical-statistical biases arise from the calculation methods of an indicator and its relation to the stochastic process that governs its motion. For example, if an indicator is conventionally measured as a three-month moving average but evolves as a random walk of monthly values, the position of the latest month versus the average predicts a forthcoming indicator change.
- Changes in information states are not information states of changes. The former are updates of perceptions that occur only on specific announcement days. The latter are point-in-time measures of changes in the economy that are available for all days, calculated based on the data “vintage” that was used on that day. Information states may change, even if the economy does not, and vice versa. For example, if a previous trend decline in economic growth has just run its course, the information state has changed, but the economic growth rate has not.
- Information state changes can be aggregated. There are two ways of aggregation: across time and across indicators. Aggregation across time is suitable if changes are frequent and information efficiency can be enhanced by tracking intertemporal averages, possibly in exponentially decaying form. Aggregation across indicators is suitable if the indicators reflect a similar concept, such as a broad inflation trend. These two types of aggregation can be gainfully combined to measure trends of information state changes of broad economic concepts. Aggregation is not trivial, however, and its basic steps are explained below.
How information state changes can be aggregated
Generally, aggregation can be performed in four steps: normalization of information state changes, a frequency adjustment, aggregation across time, and aggregation across indicators.
Normalization here means that we make information state changes comparable across different indicators and units. Normalization requires dividing changes by some interpretable metric of normal variation. For regular time series, this would be the statistical standard deviation. However, for daily changes in information states, this approach must be modified in two essential ways:
- Focus on first release dates: The variance of information states is extremely uneven across time in a predictable pattern. Variation is typically high on the first release dates for an observation period, much lower on revision dates, and zero on all other days. To have a common interpretable basis for normalization, we only consider the standard deviation of new observation period releases as the denominator.
- Expanding windows: The data used for normalization of any date must not include subsequent experiences. Otherwise, the “point-in-time” principle is violated and normalized information state changes reflect hindsight, making them less suitable for testing related trading signals. Thus, the samples used for standard deviation estimation increase over time.
The normalized change of information state with respect to a particular indicator shall be defined in units of standard deviations on new period release dates up to that date. Often, the information change series is also “winsorized” (capped and floored at a certain multiple of standard deviations) at this stage, as occasional large outliers can distort correlation analysis.
Frequency adjustment here means that information state changes are weighted in accordance with the length of the observation period. If the observation period is a quarter, for example, the inflation state change values will have a square root of three times the weight of an equivalent monthly change and approximately a square root of 13 times the weight of a weekly change. This allows aggregating across indicators of different observation period frequencies.
Temporal aggregation means summarizing normalized and frequency-adjusted information state changes over a specific lookback window. In the example below, we use simple trailing moving sums over 23 working days to replicate approximately a month’s worth of information changes. Alternatives are sums over exponentially decaying lookback windows and over a certain number of underlying observation periods.
Cross-indicator aggregation means combining different normalized, frequency-adjusted, and possibly aggregated information state series. If normalization and frequency adjustment have been performed and temporal aggregation uses the same lookback windows, this can be done by simple averaging. If there is good reason to expect one indicator to be more meaningful than others, weights can be applied, but in the below examples, only unweighted averages (“conceptual parity”) have been used.
U.S. macro information state changes
As a simple example, we track the information state changes in U.S. economic growth, inflation, and labor market dynamics, using two indicators for each concept. Also, we focused only on changes from a year ago. This is a most basic information set, but it allows for all important points related to information state changes of popular albeit volatile economic indicators. All these information types are seen as important factors of U.S. Treasury bond performance.
Changes in economic growth information: The purpose is to track normalized changes in key GDP growth trackers. Here these are trackers of estimated annual GDP growth, i.e., percent over a year ago, as 3-month moving averages, using both high-frequency indicators and the national accounts. The information states are generated by the methods of intuitive GDP growth estimation (view documentation) and technical or nowcast GDP growth estimation (view documentation). They implicitly include the information states of many U.S. activity indicators. Positive changes in growth states are expected to affect Treasury returns negatively as they support tighter monetary policy and higher real interest rates. Information changes will be put in negative terms to represent Treasury return tailwinds.
Changes in inflation information: The purpose is to track normalized changes of key inflation metrics that are watched by economists and the market these. The metrics include headline CPI inflation (% over a year ago) (view documentation) and headline PPI inflation (view documentation). Positive changes in inflation states are expected to affect Treasury returns negatively, as they raise inflation expectations and the trajectory of policy rates.
Changes in labor market information: The purpose is to track normalized changes in key metrics of labor market tightening. Here, these metrics are annual percent changes in employment (view documentation) and annual differences in (continuous) jobless claims as 4-week moving averages (view documentation). Employment growth is expected to affect bond returns negatively, while jobless growth should impact them positively. Generally, tightening labor markets supports monetary policy tightening.
The changes in these six indicators are all given the signs that turn them (theoretically) into bond market headwinds. They are normalized in accordance with the methodology described in the above section and then adjusted for frequency, which here mainly means putting jobless claims and other indicators on the same footing. Thereafter, the normalized adjusted information states are aggregated over 23-day windows as approximate monthly changes and finally aggregated pairwise to give basic aggregate growth, inflation, and employment information state changes. Finally, we aggregate the three economic concepts to a single bond macro headwind information state change.
U.S. macro information state changes and U.S. treasury returns
The target series of this analysis is the generic 10-year U.S. treasury excess return taken from the JPMaQS database (view documentation). We check the concurrent and predictive relation between the joint and individual information state changes of U.S. economic indicators and this return at a monthly frequency.
The concurrent relations between information state changes and 10-year Treasury returns at a monthly frequency have all been positive, i.e., in line with theoretical expectations. That means that downward revisions of perceptions of economic growth, inflation, and labor market tightening have been related to positive bond returns. The relations have not been very strong, however. Only the correlation between economic growth changes and Treasury returns has been significant at the 10% level. The relation between labor market tightening changes and returns has been smallest. Indeed, jobless claims growth changes did not explain bond returns positively.
The low linear correlation is not surprising, , given that the chosen indicators are only a small subset of U.S. economic data and that many other types of information pour in on that market. Moreover, as explained above, not all of these changes are surprises. There are many professional forecasters of U.S. economic data that should anticipate part of the changes in this small data set.
The relations change thoroughly if one considers subsequent monthly Treasury returns., The predictive relations between monthly economic information state changes and Treasury returns have all been negative, i.e., oppositive to the concurrent and theoretical relations. The cross-concept aggregated information state change has even been a significant predictor of subsequent bond returns.
This negative relation points to payback for overtrading. For the chosen set of high-profile U.S. indicators, exaggerated publicity and influence are plausible. Action is more newsworthy than careful economic trend analysis. U.S. growth, inflation, and labor market data are closely watched, and changes or surprises are much commented on, even though many of the related indicators can be quite unstable on a month-by-month basis. This supports the hypothesis of a cognitive bias that leads to excess positioning in the direction of the latest published changes and surprises.
This hypothesis would imply that payback is more likely, the greater the publicity and the weaker the stability of indicators that form the basis of information state changes. Stable and neglected indicators in smaller markets should not be prone to such effects.
