Saturday, March 7, 2020

Old School / New School Analysis - Feeling your way across a data stream



“Machine learning can be a very powerful tool and can be used well by those who respect markets, but unsupervised machine learning techniques that find patterns without first forming any hypothesis can also pick up spurious correlations. We retain an old school quantitative approach of forming a hypothesis and testing it. Our use of machine learning is zero.”


There is an old school versus new school battle for model building within managed futures and other hedge styles. It is a battle for hearts and minds of quantitative management and investors. Do you need a reason for a model or is successful performance enough? 

The "old school" or classic quantitative approach to investment management is steeped in the scientific method of hypothesis testing. Pose a hypothesis based on a model, and test whether it works at prediction and can forecast future prices. This old school comes from the classic foundations of statistics, econometrics, and finance. It is not about just predicting but offering explanations for why a model may work. A theory drives decision-making.

The "new school" focuses on data analysis such as machine learning and not hypothesis testing. There is a tremendous focus on successful prediction and the data will speak for itself. The emphasis is centered on practicality. Does it predict? Yes. Use it.  No. Scrap it. Data are complex and messy, and we accept our limitations at forming the right hypothesis or understanding market dynamics. If you are looking at large data sets, there may not be clarity concerning what is the cause for behavior even if there are repeatable patterns. 

Who is winning the battle between these two quant schools? If current usage is the driver of success, the new school of machine learning is dominating interest. Perhaps all of the hypotheses are exhausted. Is there a machine learning a fad? Perhaps. 

We have seen fads in techniques and hypotheses about markets. Market efficiency may have been the greatest fad on how to view market behavior. Econometric techniques have also risen and fallen in popularity only to be displaced by new ideas. However, what machine learning does have is a deeper focus on non-linear behavior which is often difficult to tease using traditional hypothesis testing. 

Nevertheless, return performance has ultimately been the driver of technique usage. If an investor does not make money with a data tool, it is thrown into the dust bin of history. Right now, the new school is still in ascent and the old school of hypothesis testing seems so conventional and yesterday. These techniques could face a form of market efficiency. As more investors use machine learning, earlier successes will diminish. 

Still, for many, there is a feeling surrounding new techniques that should err on the side of caution and conservatism. Use a new approach on old and new data, but still look for an explanation consistent with theory. Form a new hypothesis, but incorporate new data science to find proof. This does not have to be data mining but a search for deeper truth as more information is made available. 

Wednesday, March 4, 2020

The hidden fixed income benchmark risk - Is this what you want? Is this what you expect?


Grab a tiger by the tail. You have to hang on and cannot let go. This is the same as investing in the Bloomberg Barclays AGG index. One of the more important practical risk issues in fixed income concerns the changing composition of the benchmark index. See "THE PASSIVE AGGRESSIVE AGG, REVISITED Passive investing in bonds today turns prudence on its head" by Peter Chiappinelli. The AGG index has gotten significantly risker, but the problem is that it serves as the key benchmark for many investors. 

If you de-risk from the established benchmark, you will underperform in a declining rate environment. If you hold benchmark positions, you are taking on greater risk than previously held when the index had less risky attributes. Of course, no one is complaining given this duration increase is coming when rates are declining and credit spreads have been improving. Given the nice negative correlation with equities, this increase in risk has provided both diversification and safety. 

While there is a positive convexity component from longer duration when rates fall and potentially shorter duration if rates rise, there are issues with the AGG index composition. Mortgage exposure which has shorter duration and negative convexity has declined and credit risk has increased. The composition of the credit risk has moved to lower quality with BBB-rated debt. Treasury risk has also increased.



In ten years, the duration of the index has increased by over 50%.

Surprisingly, unlike their equity brothers, managers have been able to outperform the Bloomberg Barclay AGG index, but this has often been through holding even more credit and duration risk.

The reach for yield has been achieved through higher risk from just following the index. In the last week, the long duration has been a benefit, but the credit exposure has been a total return drag. The question for investor is, nevertheless, complex. Is this what you expect, and is this what you want? 

Tuesday, March 3, 2020

Conditional momentum based on the basis (carry) - It works


Trend-following and momentum strategy performance can be improved upon through combining it with carry strategies. We have described blending trend and carry at a high level as combining convex (trend/momentum) strategies with concave (carry) strategies. See Blending trend-following and carry through simple diagrams. This combination of momentum strategies (a long/short trend relative or cross-sectional time series approach) with a carry strategy, which has a better stand-alone return, can provide a better overall return pay-off.

This combination can also be integrated into a single model by making the momentum trades conditional on carry. With respect to futures trading, this combination will make the momentum trades conditional on the basis or difference between the nearby and next to nearby futures contract. The basis is positive (negative) if the market is in backwardation (contango). The conditional approach will cut short trades (negative momentum) when the basis is positive (backwardation) and cut long trades (positive momentum) when the basis is negative (contango) This a simple combined conditional model will generate a significantly improved Sharpe ratio over the unconditional approach that only uses momentum. 

The research on this comparison of strategies, which looks across a large set of markets and asset classes, is compelling. See "Carry and time-series momentum: a match made in heaven" by Molyboga, Qian, and He. Using their extensive data set, the researchers find a meaningful jump in Sharpe ratio performance. Interestingly, the large gains in the conditional Sharpe ratio comes during recessions. This is also likely the time when markets switch from backwardation to contango.



There are simple ways to improve on trend-following and momentum strategies that will generate extra return when needed during "bad times" and provide added overall return per unit of risk. This combination will cut any problems with momentum crashes and smooth returns. This strategy mix can be done efficiently in a single modeling approach. 

Monday, March 2, 2020

CTA/Managed futures managers - What is happening behind the averages?



A look at early estimates suggests that CTA / managed futures managers, on average, generated zero return for the month, yet a look at a sample of managers shows a large degree of dispersion, over 20% in a month. The median was also zero. Some managers provided protection and others matched the equity decline. The challenge for any investor is to find out which side of the dispersion continuum their choice will fall. This is not an easy task, but some questions will make allocation decision easier.
  • How many market traded? More markets offer more diversification.
  • How much equity exposure? A portfolio can be wrong-footed on an equity reversal to the downside in the short run.
  • How fast will the model adjust to different shocks? Very long trend models may not react to fast changes. 
  • Are their stop-losses? Stops offer protection from shocks.
  • Is there volatility targeting and positioning? Volatility targeting may cause delevering at the wrong time.
  • Are other strategies used beyond trend-following? This balance can provide diversification during stale time but may hurt when markets move.
These questions scratch the surface and there are not hard rules on what will happen. A strong review of shock periods using daily data is required, but last month again showed that all managed futures managers are not same.