Social Science Research Network
--> Advisor Perspectives
--> Seeking Alpha


One of the most damaging effects on your investment portfolios is economic down cycles or "recessions". Many major stock market corrections (bear markets) occur on or before recessions and reach their peak (panic selling) at the height of the recession. This is because the stock market is a forward-looking discounting mechanism on listed company earnings.

The investor who bought and held for 60 years did not have his performance impaired from recessions, but the reality is nobody buys and holds for such long periods. Invariably what happens is retail investors buy somewhere on the up-leg, most often near the top, and then a recession/correction follows and the investor lands up selling at the bottom in the panic (buy high sell low). Even if the investor had the nerves to hold on through the recession it can take several years before he lands up being even again. Consider the person who invested in May 2008 and up until September 2011 had still not recovered his bear market losses! 

Look at the chart below of the monthly JSE index plotted against recessions dated by the South African Reserve Bank (SARB) since 1960 and "major corrections/bear markets" which we have defined as at least 4 consecutive months of declines on the JSE and a fall of at least 8% :

Visually, it appears that a lot of corrections fall in recessionary periods, and this is actually the case. Consider that over the last 620 months since 1960, some 38% of these months fell into "bear market" or "corrections" depicted above, and 62% of them fell into "bull markets". If we focus on that 38% of the time the JSE is undergoing a correction (232 months in total) then exactly 52% of these (122 months) occurred during the recessionary periods shaded in grey above. Conversely, only 28% of "climbing market" periods fell into recessionary periods.  The statistics are clear - the stock market performs much better in non-recessionary periods than recessionary ones!

Taking the above paragraph further, if one could avoid the stock market during recessions, then you would avoid 52% of all corrections at the price of missing out on 28% of bull market periods. Add to this the fact that markets correct more per unit of time than they rise per unit of time (remember the old saying "The Bull climbs the staircase, but the Bear comes down the elevator") then a strategy of avoiding recessions would amount to a considerable net-gain to your investment performance.

We depict two 50-year investment strategies below - the first invests in the stock market outside of recessions (i.e. it completely manages to avoid them in a perfect world) and the other invests in the stock market only during recessions:

Now that we accept that recessions damage returns, we focus our attention to how we DETECT them starting, but also importantly know when they are ENDED (although the detection of the start is much more important than when it ends!) On the whole our Investment Timing Systems such as LBYC and SuperMODEL will get you out before the onset of recession and get you back in around the middle of the recession (when the stock markets figure the worst is over and start pricing in the recovery). However it would be nice to have a mathematical unsubjective model that gave an exact representation of the probability of recession (and hence quantified your risk). This could also supplement our timing systems and result in better timing of the markets.

In economics, a recession is a business cycle contraction, a general slowdown in economic activity. During recessions, many macroeconomic indicators vary in a similar way. Production, as measured by gross domestic product (GDP), employment, investment spending, capacity utilization, household incomes, business profits, and inflation all fall, while bankruptcies and the unemployment rate rise. Recessions generally occur when there is a widespread drop in spending, often following an adverse supply shock or the bursting of an economic bubble. Governments usually respond to recessions by adopting expansionary macroeconomic policies, such as increasing money supply, increasing government spending and decreasing taxation. The traditional measure of recession is two consecutive quarters of negative GDP growth, but this is no longer accepted in macro-economic circles. Rather, reserve banks use the definition of economic cycles defining expansions as rising economic cycles and contraction (recession) as declining economic cycles. For this reason virtually all reserve banks use composite economic indicators, which are made up of varying macro-economic components, to determine recessionary periods.

This approach was first adopted by the National Bureau of Economic Research (NBER). The NBER Business Cycle Dating Committee has been dating the U.S. expansions and recessions for the past sixty years. The members of the committee reach a subjective consensus about business cycle turning points, and this decision is generally accepted as the official dating of the U.S. business cycle.

Although careful deliberations are applied to determine turning points, the NBER procedure cannot be used to monitor business cycles on a current basis. Generally, the committee meets months after a turning point (that is, the beginning or end of an economic recession) has occurred and releases a decision only when there is no doubt regarding the dating. This certainty can be achieved only by examining a substantial amount of ex post revised data. Thus, the NBER dating procedure cannot be used in real time. For example, the NBER announced only in July 2003, twenty months after the fact, that the 2001 recession had ended in November 2001! Almost universally, academics, economists, policy makers, and businesses defer to the determination by the NBER for the precise dating of a recession's onset and end. The above also applies to the South African reserve Bank (SARB) who dates SA's economic cycles.

We have two probability models we have built to identify recessions. The first uses the SARB Leading Economic Indicator KPB7090N (LEI) and the 2nd one, which focuses on detection of U.S recessions, is a 4-factor model that uses the Philadelphia Federal Reserve (Filly Fed) Business Outlook Survey, the sophisticated Aruoba-Diebold-Scotti (ADS) real-time business conditions index (also maintained by the Filly Fed), the Conference Boards' Employment Trends Index (ETI) and the truly excellent e-Forecasting US real-time Leading Economic Indicator (e-LEI).

The U.S based model is the most sophisticated, accurate and probably the most important for us. When the US sneezes, everyone catches a cold. Also, US recessions lead SA (and most country) recessions by 2-6 months and so the US model is our "early warning system" or "Canary in a coal mine." In South Africa we just do not have similar sophisticated surveys, business condition indexes or "real-time" leading economic indicators going back 60 years (so we can develop robust models) so in effect we rely heavily on the US model to warn of local recession.

The SA model is updated once per month (normally end of the month) for our subscribers, whilst the U.S model is updated once per week as underlying data becomes available. A comprehensive analysis of all the models is published in The Monthly JSE Pulse (MJP)

Our probability models for SA recession are based on the observed behaviour of the SARB KPB7090N Composite Leading Indicator. This indicator consists of 11 economic metrics that have leading characteristics (i.e. their observed status gives a future view on the SA economy) such as manufacturing, retail sales, job adverts in Sunday Times, money supply, commodity prices etc. There are 4 different models we use with KPB7090N, namely a Real-observed vector recognition model, a Polynomial probability distribution pattern recognition model, a Linear distribution pattern recognition model and a Probit Logistic single-factor statistical regression model.

We then plot a probability range that shows the highest & lowest reading from the 4 models for that month. The probability range is shown by the shaded green area in the chart below. The range shows the probability of recession within 0-3 months and reads from zero (improbable) to 100 (a certainty). You can see on the right-hand edge it is showing a probability range of 11.11% to 28.6% for the month of August 2011. The range is depicted in green lettering on top right of the chart:

When the probability range reaches into the 50%-60% red shaded area there is a 90% confidence level that we are already in or will be in a recession in 0-3 months time. Of the 11 times this has occurred over 50 years, 10 signalled a genuine recession (one false positive.)

The 50-year track record of all 4 SA models are shown below against SARB determined recession dates. For those statistically inclined, they all have Pearson R-squared values of 0.73 to 0.75 meaning the models are all statistically significant (good fit to SARB recession dates).

The Probit model has the highest R-squared and also only 1 false positive versus the other model's 2 false positives. While we are on the false positive topic, even though SARB elected not to classify these periods (where there were false positives) as recessions, the spikes in probability meant some stress was occurring in the leading economic index. As such, virtually every spike in these probability models is accompanied by stock market corrections, regardless of whether a recession occurs or not. This is important to note as it means these probability models give advance warning of corrections!

It is important to combine the SA recession model with that of the U.S model when making investment assessments. Economic time-series data for the US is far more robust than that in SA, going back to 1920's in some instances. For this reason we can build far more robust models for the US than for SA, where the data is hard to come by and only goes back 2 or 3 decades. As a result our U.S models have r-squares upward of 0.9 as opposed to the SA' models' 0.75. We discuss the US models below.

U.S recessions lead SA recessions by 2-4 months. That is because the US is the worlds consumer and we are providers to them and the countries that manufacture goods for them. Their economy slows down before it impacts ours. We are therefore very interested in the probabilities of US recession to act as "canary in the coal mine" for SA recessions, and to complement our less accurate SA recession probability models.

We work with four composite  base-indicators (as opposed to just KPB7090N with the SA model) namely the Philadelphia Federal Reserve (Filly Fed) Business Outlook Survey, the Aruoba-Diebold-Scotti Business Conditions Index (ADS),the Conference Board Employment Trends Index (ETI) and finally the e-Forecasting Leading Economic Index (e-LEI). Each of these on their own are good recession forecasters, but when combined into a four-factor multivariate Probit logistic regression model, they are very powerful indeed, yielding a model with a Pearson R-squared of 0.90 (implying a very strong fit to NBER recession dates). The combined model is unique in that it is taking into account about 29 discreet economic variables across all 4 base indicators to arrive at a probability of recession.

1.The Filly Fed Business Outlook Survey
The Business Outlook Survey is a monthly survey of manufacturers in the Third Federal Reserve District. Participants indicate the direction of change in overall business activity and in the various measures of activity at their plants: employment, working hours, new and unfilled orders, shipments, inventories, delivery times, prices paid, and prices received. The survey has been conducted each month since May 1968. We use the surveys' "current activity" diffusion index (shown below) for recession forecasting:

A logistic regression statistical model for recession probability forecasting based on the current activity diffusion index appears below. Whilst it appears to pick out NBER recession dates nicely it is volatile and thus prone to false positives (false alarms). As of 30 Aug 2011 it is showing an alarming 87% probability of recession in the U.S within 0-3 months. However you can see it is very volatile with many false positives and only has an R-squared of 0.67. This model has the lowest weighting in the multi-factor model due to the high number of false positives.

2.The Aruoba-Diebold-Scotti (ADS) Business Conditions Index
The ADS Business Conditions Index is designed to track real business conditions at high frequency. Its underlying (seasonally adjusted) economic indicators (weekly initial jobless claims; monthly payroll employment, industrial production, personal income less transfer payments, manufacturing and trade sales; and quarterly real GDP) blend high- and low-frequency information and stock and flow data. The ADS index is updated as data on the index's underlying components are released.
The ADS business conditions index is based on the framework developed in Aruoba, S.B., Diebold, F.X. and Scotti, C. (2009), "Real-Time Measurement of Business Conditions," Journal of Business and Economic Statistics 27:4 (October 2009), pp. 417-27.

The statistical prediction model we built for the ADS is shown below. We note it has very high R-squared values (statistical significance of the model output) and at the 50% level (lower red horizontal line) we have 2 false positives and at the 60% level we only have one false positive. Also note how the probability curve nicely encapsulates the NBER recessions, very accurately  falling rapidly off as the recession ends. Just think about it - this model is giving near perfect recession dating some 6-12 months before NBER proclaim recession dates! The model has a very high R-squared correlation of 0.85 and has the highest weighting in the multi-factor model.

3.The Conference Board Employment Trends Index (ETI)
The Employment Trends Index (ETI)™ offers a short-term, forward look at employment on its own. It gives economists and investors a new forecasting tool. It also helps business executives sharpen their short- to medium-term hiring and compensation planning. The Employment Trends Index aggregates eight labour-market indicators, each of which has proven accurate in its own area. Aggregating individual indicators into a composite index filters out so-called “noise” to show underlying trends more clearly.

The eight labour-market indicators aggregated into the Employment Trends Index include:

1.Percentage of Respondents Who Say They Find “Jobs Hard to Get” (The Conference Board Consumer Confidence Survey),

2.Initial Claims for Unemployment Insurance (U.S. Department of Labour),

3.Percentage of Firms With Positions Not Able to Fill Right Now (National Federation of Independent Business Research Foundation),

4.Number of Employees Hired by the Temporary-Help Industry (U.S. Bureau of Labour Statistics),

5.Part-Time Workers for Economic Reasons (U.S. Bureau of Labour Statistics),

6.Job Openings (U.S. Bureau of Labour Statistics),

7.Industrial Production (Federal Reserve Board),

8.Real Manufacturing and Trade Sales (U.S. Bureau of Economic Analysis).

The Conference Board publishes the Employment Trends Index monthly, at 10 a.m. ET on the Monday that follows each Friday release of the Bureau of Labour Statistics employment situation report. For more details on the index you can go look at the Conference Board Product Description.

A chart of the ETI which has been published since 1973 and overlaid against NBER dated recessions appears below. Generally speaking it would appear that when the ETI starts declining, a recession follows soon after. It also appears that when the ETI rises after a decline, it is a great early warning to the end of recession.

The statistical non-linear Probit regression prediction model we built for recession forecasting with the ETI is shown below. It uses the 3-month %change in the ETI as input.  It is prone to less volatility than the other two models, but it suffers from 2 false positives at the 0.4 probability level. It has the 3rd highest weighting in the multi-factor recession probability model.

4.The eForecasting Leading US Economic Index(eLEI)
There are 3 reliable "subscription paid" leading economic composite business indexes for the US, namely the Conference Boards' LEI, the ECRI weekly indicator and the e-LEI from eForecasting. We have tested recession probability models on all of their data going back to 1960 (7 recessions) and the eLEI and the LEI offer similar performance and both out-perform the proprietary ECRI index.

However, the eLEI is much more timely than the LEI, and presents data with 1-month lag as opposed to the LEI which has a 2-month lag (most country LEI's have 2-3 month lags.) In fact this is a rather unique factor of the eLEI over all other monthly leading indicators. This can make a significant difference to forecasting timeliness. Another important aspect of the eLEI is that it is not subject to revisions which provides for far better forecasting accuracy.

The eLEI uses 7 leading components of the US economy, namely :

1.Manufacturers' new orders
2.US Stock market prices
3.Consumer expectations
4.Housing activity
5.Interest rate spread
6.Manufacturers productivity
7.Foreign Demand (another very unique aspect of the eLEI)

The historic values since 1959 of the eLEI index and NBER-dated recessions are shown below:

Using the eLEI we have built a non-linear Probit statistical regression model for U.S recession probability measurement as shown below. Since the eLEI is a leading indicator, it generally gives us better advance warning of recession than the other indicators (looking below you can see it was never "late" in calling recession although it did suffer from two false positives. Still, 8 right calls out of 10 gives us 80% accuracy.)

The history for eLEI is longer than all the other indicators we have tested and as you can see since 1959 it suffered from only 2 false positives at the 0.3 level which is the lowest false-positive threshold of all 4 the models.

5. The Combined 4-factor model
This uses all 4 the previous composite indexes as input into a 4-factor non-linear logistic regression equation to yield the below recession probability model. The threshold at which we have zero false positives (false alarms) is now at 60%, lower than all the single-factor models. More importantly, the threshold at which we only have 1 false positive is a low 0.2 (20%), again better than any of the individual models. We are also at that magical 0.9 R-squared meaning the model has a 90% fit to NBER recession dating. Not only does the multi-factor model provide for lower false positive thresholds, but it also has higher R-squared correlations than any of the single factor models.

Note that from 1968 through to 1973 we used just a 3-factor model as no ETI data was available during this time frame. Also note how much less volatile the readings are in the model for periods where no recession were present - the model hugs zero for most of the time, only rearing its head when large amounts of stress start being encountered by the economy. Have a look at this chart and then scroll your screen back to the first sets of charts we displayed to see how magnificently the multi-factor model reduces volatility of the probability measurement.

6. Probability gauges
We publish recession probability gauges in addition to the probability charts for our subscribers on a weekly basis in The JBAR Report as shown below. The gauge on the top "QUATTRO" is the 4-factor model and the sub-components are shown below it.

The coloured calibrations (green, yellow and red) are set according to the appropriate thresholds taken from each models' historical probability chart. When the needle is in the green, there is nothing to worry about. When the needle is in the yellow we need to be on alert. When the needle is in the red it is a virtual certainty that we are already in a recession or will be in one within 1-3 months time.

7. New enhancements to the U.S model (OCTOBER 2011)
The U.S Recession Dating & Probability model discussed above has been improved by our research team in October 2011 to include 3 extra monthly-published factors for better accuracy and reliability:

1. The e-Forecasting monthly GDP estimates (eGDP)
2. The Chicago Fed National Activity Index (CFNAI)
3. The Institute for Supply Management (ISM) manufacturing survey (ISM-PMI)

They are incorporated just as the other 4 factors discussed previously, into a composite 7-factor model (called RPM7) as shown below for 1 November 2011 (click on the chart for a larger version)

Whilst this model has a slightly improved r-squared, its main benefit comes from more timely flagging of recession starts and ends, and zero false positives (no false alarms) at the 0.5 (50%) probability level. At the 0.3 (30%) probability level it produced only 1 false alarm in 2003 where technically we feel the SARB erred in not flagging as a recession as the economy was under severe contractionary forces at the time. Believe me, it felt like recession even if SARB elected not to call it one!

One of the improvements is that we use the parameters of the probability model to produce an "Economic Output Index" or EOI, as shown by the green line. It is essentially that aggregation of the underlying 7 factors that produces the optimum recession dating model and therefore serves as an ideal representation of the state of the US economy. When it dips below zero, we are in contraction (recession) and when it is above zero we are in expansion. Not only is this very useful to get a feel for the power and direction and trend of the US economy, but it provides us with directional context not available from the probability model. For example we can see we are on a down-trend since May 2010 and are now bordering on contraction.

There is a simple rule used by the model for recession dating, namely call a recession when the probability rises above 30% and call the end of recession when it falls below 15% (green line). Alternatively call recession when the EOI falls below -5 and its end when it rises above -5. These rules on average call starts 0.42 months after official dates and the ends 0.42 months after official dates. Given that official dates are only proclaimed by NBER up to 8-12 months after the event, that is a pretty decent dating system, useful and timely enough for investors to use.

Another new feature we introduced is the RECESSION DIFFUSION INDEX. For each of the 7 underlying probability models, we calculate that "Golden ratio" which is the threshold value for which when the calculated probability rises above it to call a recession, and falls below it to call the end to recession, the model  (a) makes the least amount of false positives (false alarms) and (b) the least amount of false negatives (not calling a recession when there was one already in place.) These thresholds are the optimal values for dating recession for each model. We than track an index showing how many of the 7 models are above their "Golden ratios" (i.e calling for a recession) and call it the DIFFUSION INDEX shown below as per the right-hand scale of the chart:

The solid purple line shows how many of the 7 individual single-factor probability models are calling a recession, whilst the solid red line shows the simple rules described in the previous paragraph of the 7-factor model in action in the 2008 recession. The DIFFUSION is a useful decision support tool to allow us to further assess the risk of recession together with the simple rules. Whenever at least 4 models are flagging recession simultaneously, this can give you a very accurate advance warning over and above the signal from the 7-factor model.

8.Probability Update Schedule
The components of the 7-factor Recession Probability Model (RPM7) are updated at varying time frames and frequencies which means that during the course of any month, RPM7 may be updated on up to 18 separate occasions. As a result, it acts as a REAL-TIME reflection of the US Economy and the probability of recession.

For example, the ADS is updated weekly as weekly jobless claims are published and is also updated 3-4 times per month as various monthly economic data such as industrial production etc. become available. The BCS is updated once per month on or around the 15th of each month. The ETI is updated in the 1st or 2nd week of the month (for the previous month), and the eLEI is updated on the first business day of each month. Below is a sample of a "update schedule" for the recession probability model as at 3rd November 2011:
The lighter shaded items have already been published and incorporated into RPM7 and the darker ones are in the future. Individual weeks in the month are demarcated by thin borders (5 weeks are shown in the above example.) Each item is flagged with a GOOD or BAD depending of it got worse or got better over the previous month.

RPM7 is published weekly for subscribers and incorporated into a dedicated sheet in the weekly JBAR report published on Sundays:

We update and analyse RPM7 every day, so if something material should occur in-between weekly reports, subscribers will be notified by an email alert.


--> Social Science Research Network
--> Advisor Perspectives
--> Seeking Alpha

Make a Free Website with Yola.