Macro information state changes are point-in-time updates of recorded economic developments. They can refer to a specific indicator or a broad development, such as growth or inflation. The broader the economic concept, the higher the frequency of changes. Information state changes are valuable trading indicators. They provide daily or weekly signals and naturally thrive in periods of underestimated escalatory economic change, adding a layer of tail risk protection.
This post illustrates the application of information state changes to interest rate swap trading across developed and emerging markets, focusing on six broad macro developments: economic growth, sentiment, labour markets, inflation, and financing conditions. For trading, we introduce the concept of normalized information state changes that are comparable across economic groups and countries and, hence, can be aggregated to local and global signals. The predictive power of aggregate information state changes has been strong, with material and consistent PnL generation over the past 25 years.
The post below is based on Macrosynergy’s proprietary research.
Please quote as “de la Porte Simonsen, Lasse, Sueppel, Ralph and Tyagi, Palash, ‘Macro information changes as systematic fixed income signals,’ Macrosynergy research post, September 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 for macro-quantamental indicators. J.P. Morgan offers free trials for institutional clients. Access to JPMaQS for research alone (excluding the latest months of data) is generally free.
Also, an academic research support program sponsors data sets for relevant projects.
The basics of macro information state changes
A macro information state is the latest instance of an economic indicator available to the public on a given date. For example, on September 18, 2024, the latest annual core CPI inflation rate of Japan was reported as 2.7% based on a change of the CPI for July 2024 versus the CPI for July 2023. The timestamp of an information state, here September 18, is different from the observation period of the economic event, here the period July 2023 to July 2024. Time series of information states represent the evolution of public records of recent economic developments. They are point-in-time data, like market prices, and unlike standard economic data. See also a previous post on the subject here.
Macro information state changes are simply the first differences of recorded information states. Conceptually, they are updates of publicly available indicators. For example, on September 19, the latest Japanese core inflation rate was updated to 2.8%, now referring to the period August 2023 to August 2024. The information state change on that date was 0.1%. Information state changes result not only from changes in observation periods but also from revisions of previously released data, changes in model parameters, such as calendar factors, or changes in reference series of an economic concept, such as the CPI data that a central bank uses to estimate core inflation.
Information state changes often refer to broad economic concepts rather than single indicators. For example, information states of broad inflation pressure in a country should consider various price indices and changes thereof. Information states of global inflation pressure should consider various metrics of price growth across a range of larger economies. The broader the concept, the more indicators become relevant and the higher the frequency of the information state changes. Information state changes in broad economic growth and inflation are almost daily.
Information state changes are not quite the same as surprises. Surprises would be information state changes that are not anticipated by the market. In practice, information state changes are partly predictable, for example, through early indicators and autocorrelation. Broad information state changes are often autocorrelated, i.e., past changes indicate future changes, as many indicators are subject to similar background trends. But predictability does not mean irrelevance. Whilst an increase in inflation may be predicted by a conscientious economist, many market participants and policymakers respond to facts rather than forecasts or prefer to save information costs. Generally, information state changes and surprises are different concepts, each relevant to markets.
The benefits of information state changes as trading signals
Information state changes are a precious type of trading signal that can be derived from point-in-time economic information:
- Higher-frequency signals: Information state changes provide higher-frequency macroeconomic trading signals than standard economic indicators. This is because information states are more variable than actual economic trends. Economic data are subject to all types of transient influences, such as unseasonal weather, holiday patterns, large-ticket transactions, and reporting changes. The disparity between actual economic change and information state changes is particularly pronounced for broad concepts. While global inflation cycles may be measured in years and decades, many specific price indicators experience idiosyncratic cycles and short-term trends or pick up relative price changes. Short-term information state changes matter even if they do not correctly predict a shift in longer-term trends because they still shape the subsequent flow of news and views expressed by market analysts. Most short-term information state changes are, with hindsight, considered noise by economists. However, for traders, they can be profitable noise.
- Macro risk protection: Occasionally, economic change is unanticipated and highly consequential. This includes periods of self-reinforcing escalating economic dynamics, such as recessions and spiralling inflation. Signals based on information state changes typically keep positions on the side of underestimated trends as they follow the change in economic information swiftly and dispassionately. This is opposite to convention in the economic profession, where analysts often “fight for their views” and adjust predictions sluggishly in accordance with publication schedules. Most information state changes may be “false dawns” whose influence on the market is small and fleeting. However, when an unanticipated change escalates, it can enforce a large-scale repricing of expectations and risk. In the below empirical analysis, fixed income signals based on information state changes did a particularly good job in periods of escalatory developments.
In the past, data on economic information states had been hard to procure and work with. However, now they can simply be downloaded from the J.P. Morgan – Macrosynergy Quantamental System (JPMaQS). JPMaQS provides daily end-of-day (New York time) information states of major macroeconomic indicators, called “quantamental indicators”, for up to 40 countries and typically over 2-4 decades, based on concurrent data “vintages.” For research purposes, access to these indicators (excluding the last few months of data) is free.
Example: relevant information state changes for rates markets
Here, we show the relevance of macro information state changes for global interest rate swap markets. Altogether, we consider five category groups for 22 developed and emerging countries:
Economic growth: These are information states of indicators representing GDP growth and aggregate demand growth in the economy. They include the following quantamental categories:
- Intuitive real GDP growth, i.e., the latest estimable GDP growth trend based on actual national accounts and monthly activity data, based on sets of regressions that replicate conventional charting methods in markets, % over a year ago (view documentation).
- Technical real GDP growth, i.e., the latest estimable GDP growth trend based on actual national accounts and monthly activity data based on supervised learning and standard nowcasting: % over a year ago (view documentation).
- Industrial production, adjusted for seasonal effects, working days and holidays: % of the latest 3 months over the previous 3 months at an annualized rate (view documentation).
- Real private consumption, % over a year ago, 3-month average or quarterly (view documentation).
- Real retail sales, % over a year ago, monthly or quarterly frequency (view documentation).
- Merchandise imports in local currency, seasonally adjusted, % of the latest 6 months over the previous 6 months at an annualized rate, monthly frequency (view documentation).
Economic sentiment: These include information states of survey-based confidence scores of various areas of the economy, namely of the following categories:
- Manufacturing business confidence, seasonally adjusted normalized (z-)score (view documentation).
- Services business confidence, seasonally adjusted normalized (z-)score (view documentation).
- Construction business confidence, seasonally adjusted normalized (z-)score (view documentation).
- Consumer confidence, seasonally adjusted normalized (z-)score (view documentation).
Labour market: This looks at information states with respect to unemployment rates and employment growth rates:
- Unemployment rate, seasonally adjusted, 3-month moving average (view documentation).
- Employment growth, main local measure, % over a year ago, monthly or quarterly frequency (view documentation).
Inflation: These are information states of various consumer and producer price growth rates that market participants and central banks conventionally monitor:
- Headline CPI inflation, % over a year ago, including early estimates where available (view documentation here and here).
- Core CPI inflation, % over a year ago, local standards, including early estimates where available (view documentation here and here).
- Producer price inflation, % over a year ago, local standards (view documentation).
Financing: The final group contains information states of credit and liquidity growth rates as approximations of ease of financing:
- Private credit growth, change over 12 months ago, as % of GDP, seasonally and jump-adjusted (view documentation).
- Narrow money, seasonally- and jump-adjusted, % change over a year ago (view documentation).
- Broad money, seasonally- and jump-adjusted, % change over a year ago (view documentation).
- Net international investment position, as % of GDP (view documentation).
- International liabilities, as % of GDP (view documentation).
The currency areas for which these series are gathered are nine developed markets and 13 emerging markets. The developed market currency areas are (in alphabetical order if the currency symbol): Australia (AUD), Canada (CAD), Switzerland (CHF), the euro area (EUR), the United Kingdom (GBP), Japan (JPY), Norway (NOK), New Zealand (NZD), Sweden (SEK), and the United States (USD). The emerging markets are Chile (CLP), Colombia (COP), Czechia (CZK), Hungary (HUF), (ILS), India (INR), South Korea (KRW), Mexico (MXN), Poland (PLN), Thailand (THB), Turkey (TRY), Taiwan (TWD), and South Africa (ZAR). Periods of illiquidity, such as the early 2000s in some emerging markets and Turkey after 2020, have been excluded.
Standardizing information state changes
Generally, we define information state changes for specific economic indicators as normalized daily changes in units of standard annualized changes of that indicator. The purpose of normalization is to make information state changes comparable across different indicators and units, thus allowing aggregation.
For regular time series, normalization would just mean division by daily standard deviations. However, the daily variance of information states is extremely uneven across time. 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.
Furthermore, we need to “annualize” the changes to make high and low-frequency economic reports comparable. The change of a quarterly report conveys more information than the change of a weekly report, as more time has passed. Hence, we divide by the square root of the number of observation periods in the year, assuming that those changes are roughly approximated by an unbiased discrete random walk. We call the output of these transformations normalized information state changes in annualized units (NICA). Typically, these normalized metrics are winsorized to de-emphasize extreme outliers.
Note that in this post, all information state changes are plotted and applied with the sign of their expected impact on duration positions. Thus, for inflation and growth changes, a negative sign is applied. The below chart shows normalized information state changes for a U.S. GDP nowcast that is based on a range of activity indicators. This time series looks very different than standard economic indicators related to growth, with volatility dominating trends most of the time.
Aggregation of information states
The value proposition of information state changes as a trading signal relies on the principle of rational information inefficiency, i.e., the choice of investors not to follow and act on each potentially relevant new data release in real-time (view post here). Rather, many investors limit themselves to reviewing economic changes on a periodic basis, maybe weekly, relying on aggregation across economic data and countries, often filtered by economist briefings and news reports. Of course, there are some key data that are widely received in real-time, such as the U.S. labour market report. However, the full range of relevant economic data changes is too vast for a human investor to analyze in real-time and even challenging to follow periodically.
Therefore, this post proposes to model and predict gradual information updates based on standardized information state changes. This typically requires various types of aggregation.
Temporal aggregation: It is plausible that new information permeates the markets over a couple of days. An efficient version of that process can be modelled through exponential moving averages of past normalized information state changes. We have chosen a half-life of 3 days and separately tested 5 days, with similar results. Broadly speaking, this assumes that information is being mostly absorbed over the course of a week or two, which corresponds with the frequency of many research reports and briefings.
Cross-indicator aggregation: What matters for market perception is not so much the change of a single indicator as a change in a broad trend or state, such as economic growth or inflation. For example, if a range of indicators related to aggregate demand and output all show an acceleration, the message has greater publicity and credibility and, hence, probably greater impact. Here, we aggregate by forming linear averages of normalized information state changes across all indicators that belong, respectively, the groups of economic growth, economic sentiment, labour market, inflation and financial conditions. The averages are unweighted for simplicity, but weights could also be set based on judgment of data importance or optimized through statistical learning.
Cross-group aggregation: To arrive at a single plausible diversified trading signal, one can aggregate conceptual group averages of normalized information state changes as a form of “conceptual parity”. This is possible through linear combination because the scales of group values are comparable, and their signs are all set such the positive values are presumed to have a positive effect on duration returns. As for cross-indicator aggregation, this post uses unweighted averages, but optimized weightings based on past explanatory power and statistical learning would be an option (view methodological post here).
Global aggregation: Finally, country-specific information state change metrics can be aggregated into global measures. This is appropriate for directional trading across connected markets and economies, where trends in one country matter for others. Again, a linear combination is suitable as an approximation. However, given the different sizes of economies and their financial markets, plausible weights must be applied. For this post, we used point-in-time values of the shares of local GDP in world GDP in USD terms, based on a 3-year moving average (view documentation).
Aggregation naturally increases autocorrelation of information state change indicator. Temporal aggregation introduces serial relations mathematically. Cross-indicator, cross-group, and cross-country aggregation effectively combine related economic phenomena, and to the extent that common background factors drive them, they often change in the same direction. Still, as long as the initial temporal aggregation uses short horizons, aggregate information state changes are still short-term signals by the standards of economic factors, as they typically change direction every few days.
The PnL value of aggregate information state changes
We test the predictive power and economic value of aggregate information state changes for the overall economy and various conceptual groups. This follows a general evaluation method that considers the seasonality and diversity of macro trading signals across countries (view post here).
The below facet of scatterplots and regression lines visualizes and quantifies the relation between information state changes at the end of a week and the subsequent weekly duration returns across the full panel of 22 countries for the near 25-year period 2000 – 2024 (September). The critical metric is the significance of the weekly forward correlation, which is based on a special panel test that adjusts the data of the predictive regression for common global influences across countries (view post here). This test assumes that the relation between features and targets is similar across countries and that the country-specific features matter, not just their global averages.
The panel regression analysis shows the positive predictive power of aggregate short-term information state changes for all conceptual economic groups and the aggregate. The changes related to composite and economic growth have been most significant, with probabilities of non-accidental relationships of over 99%. The probability of significance in the other four groups has been between 86% and 92%. Beyond, the predictive power of the composite information state change has also been highly significant at a daily or monthly frequency.
The daily accuracy of information state changes with respect to the next day’s returns has been above 50%. Also, the parametric and non-parametric forward correlation has been positive and highly significant for all macro groups except financial conditions.
Finally, we can measure the economic value of the information state changes as trading signals based on naïve PnL metrics. Naïve profit and loss series can be calculated by taking positions by taking one unit of expected volatility per unit of normalized signal. Positions are adjusted daily, using the previous day’s information state change values and adding one day of slippage of trading into the new position. The trading signals are capped at a maximum of two standard deviations as a reasonable risk limit. The naïve PnL does not consider transaction costs or risk management rules. It will thus overstate the economic return for large amounts of assets under management but gives an objective evaluation of the signal value itself. We consider a market-neutral strategy and a long-biased strategy, whereby the latter always adds one standard deviation of the information state change signal. All PnLs are scaled (not volatility targeted) to 10% annualized standard deviations for joint graphical representation.
The 25-year risk-adjusted return of a strategy based on the composite information state changes across 22 currency areas has been material. The long-term Sharpe ratio of the naïve PnL has been 1.4 with near zero correlation to the 10-year Treasury return and without counting compounding effects. The Sortino ratio has been as high as 2.2. Seasonality has been modest for a macro signal, with consistent value generation across decades and only the 2009-10 episode of sustained negative PnLs. The PnL contribution of the 5% best-performing months has been less than 50%. Meanwhile, the information state changes provided particularly high value in times of escalating economic change, such as around the great financial crisis or the COVID-19 pandemic and the subsequent reflation period.
A long-biased strategy would still have produced a 25-year Sharpe ratio of over 1, compared to a Sharpe ratio of 0.4 for a simple long-only risk-parity exposure across all 22 markets. Thus, the information state change has also proven its value as a cross-market and intertemporal risk allocation overlay signal.
Naïve PnL generation has also been positive for all individual economic groups’ information state changes. Each of them would have added significant value. This testifies to the broad and pervasive predictive power of macro information changes. The most “profitable” field of information state changes has been inflation, albeit sentiment-based signals have delivered greater consistency. In general, the seasonality of individual economic concepts has been greater than for the composite signals, highlighting the benefits of diversification. Different economic trends matter in different periods. The only exception has been survey changes, which have generated PnL values almost linearly across time.
An alternative to country-specific signals is combined local-global information state changes. Rather than just considering the information on the local economy, these signals consider local and global changes equally. In practice, participants in local fixed-income markets follow international developments. Those in smaller economies often pay more attention to U.S. or euro area data than to their own economic reports.
Indeed, the predictive power of combined local and global information state changes has reached even greater significance than local information alone. All macro group changes have displayed a weekly forward correlation with a probability of significance of at least 93%. Also, the daily accuracy of return predictions has been higher than for local signals alone.
Considering both local and global information state changes has produced slightly better risk-adjusted returns in naïve PnLs without greatly changing the overall return profile. The 25-year Sharpe ratio of the unbiased signals would have been 1.5, and the Sortino ratio 2.5. The Sharpe ratio for the long-biased signal would have been 1.3, and its Sortino ratio 1.9.
