The hidden impact of changing market conditions on systematic bond trend models
As quantitative investment professionals we rely on data to drive investment decisions and generate performance or take risk exposures for our clients. Therefore it is normal to analyse past data in order to design investment models. We call this process backtesting: it allows us to simulate an investment strategy using historical data and use this to analyse risks, performances, drawdowns, and the behaviour of systematic strategies with respect to changing market conditions and parameterisations. Furthermore, it also helps us assess a systematic strategy’s rationale. Confidence in the rationale and behaviour of such strategies may eventually lead to capital commitments and investments. In all this process, which starts from idea generation and eventually ends with capital commitment, data is of paramount importance: it allows us to analyse and understand how an investment strategy reacts in the real world. However with experience we have recognised that data is only a part of the story, as it is backward and not forward looking.
“You can drive a car by looking in the rear view mirror as long as nothing is ahead of you.”
Bill Joy, computer engineer, co-founder of Sun Microsystems
It is the understanding of an investment strategy – its assumptions, risks and performance drivers – that is important to us. That is why we, at GAM Systematic, continuously monitor our investment models, even the successful ones, and continue to review and adapt them to changing market conditions. Here we want to discuss how shifts in market paradigms can have far reaching impacts on the way models are constructed. In particular we present the example of how changing market conditions, such as shifts in monetary policy, can impact portfolio construction methodologies and model dynamics by considering the case of trend models in the bond market.
Trend models are a type of investment strategy which buys (or ‘goes long’) assets with recent positive price trends and sells (short) assets with recent negative price trends. As such they seek to exploit the persistence of returns in given designated markets. Such strategies are also known as time-series momentum or trend following and have been widely studied by practitioners and academics1. CTA, managed futures and systematic global macro hedge funds seek to exploit priced-based, trend-following algorithms to trade futures contracts. As shown in Exhibit 1 trend models are generally constructed around four building blocks: investable universe, signal generation, portfolio construction and risk management.
Exhibit 1: The four pillars of a systematic trend following strategy
In particular, bond trend models tend to allocate to bond futures across multiple durations (bond maturities which contribute to interest-rate sensitivity) and markets. Typical investment universes include the contracts reported in Exhibit 2. In order to identify the trend, practitioners rely on a multitude of signals, from simple moving averages2 to Bollinger bands3 and / or to linear temporal regressions4. For instance, several methodologies used by practitioners in order to extract trend signals are reviewed in academic literature5. A common portfolio construction methodology relies on allocating according to the inverse of volatility (a measure of the degree of variation of asset returns), ie contracts with lower volatilities warrant higher allocations. The sum of the weights is sometimes also normalised to a suitable gross exposure.
Exhibit 2: Typical investment universe of bond trend following strategies, spanning across different regions and bond maturities
In this framework, each position is sized to target the same amount of volatility, both to provide diversification and to limit the portfolio risk from any one market. But other approaches are also used: they span from equal weights to more involved schemes where allocations are based on combinations of signal strength and volatility. Finally, risk management ensures that leverage, overall position sizing and ex-ante risk are managed according to the investment objectives and guidelines.
The inverse of volatility is widely used among practitioners and has demonstrated to be a very successful portfolio construction methodology: robust, simple and reliable. The kind of methodology we like: does a good job without any superfluous complexity. But, as we shall see, changing market conditions have called for its revision. This is where backward and forward looking investment choices come into play: without trying to sound too much like a compliance officer, it is worth reiterating that past data is not always a good indicator of current and / or future outcomes. We firmly believe it is essential to use a backward and forward looking approach to understand model dynamics and implications, and finally to drive investment decisions.
In the last few years central banks around the world have been cutting interest rates as part of their measures to fight deflationary scenarios. Eventually they reached record lows and, in some cases, negative levels. Negative interest rates are an unconventional monetary policy tool, but in today’s world they have become normal. The Swiss government ran a de facto negative interest rate regime in the early 1970s, by charging a penalty on increases in Swiss franc cash deposits of non-residents. This was the answer to Swiss franc appreciation due to investors fleeing inflation in other parts of the world. But after the financial crisis of 2008 very low interest rates appeared in all the developed economies, in order to boost growth and fight deflationary pressure. In July 2009, negative interest rates were deployed by Sweden's central bank when it cut the overnight deposit rate to -0.25%. The European Central Bank (ECB) followed in June 2014 when it lowered its deposit rate to -0.1%. Other European countries (ie Switzerland and Denmark) and Japan have since chosen to pursue negative interest rates. The Federal Reserve’s target rate did not actually reach negative values but hovered around zero for a number of years.
But why and how is monetary policy influencing portfolio construction methodology? There are at least two reasons / connections. On the one hand, the statistical properties of bond time series have changed in the last few years; on the other, the behaviour of the term structure of interest rate is also structurally different. To us, this means the portfolio construction methodology, and, the signal generation, should be questioned and reviewed.
Bond time series have changed because at very low interest rate levels further significant decreases are less likely. Even if one could argue that this level was zero a few years back and negative today, market expectations set a lower limit to interest rate levels. As we approach this limit the return distribution becomes skewed (Exhibit 3).
Exhibit 3: The distribution of interest rate movements might become highly asymmetric when the lower expected boundary is approached. At the same time volatility decreases as the distribution is more peaked
To understand this point, let us assume that interest rates are zero and they cannot go below that level. As a matter of fact, in our example, they would either go up or stay the same, hence inducing an asymmetry in the distribution. The situation is very different when interest rates are at 3% and have the potential to move either up and down. In such a case of asymmetric distribution, we believe volatility is simply no longer a good measure of risk. Moreover as the distribution becomes very peaked, the realised volatility decreases, possibly introducing a distortion in the portfolio weights. Essentially, asymmetric contracts tend to be afforded higher weights potentially altering the behaviour of the strategy. Depending on the portfolio construction methodology and on the investable universe the impacts could be large. These new market conditions force us to question and closely monitor the use of volatility in the model, either when employed in the signal generation or in the definition of the weighting scheme.
Exhibit 4: Yield curve for US treasury bonds at end December 2016, showing the typical upward sloping structure
The term structure of interest rates shows the values of rates on the same assets with different maturities. The term structure is displayed in what is known as a yield curve, an example of which is reported in Exhibit 4. While it is typically upward sloping, the yield curve shifts and the gradient adjusts in response to economic changes. It is also an indicator of different phases of the economy over time and historically it has indeed been used for that purpose – as noted in a raft of academic research. In ‘normal’ times central banks set short-term rates in order to control the supply of money to the economy6. During the recent financial crisis, a form of monetary policy, known as quantitative easing (QE), was used to further increase the money supply. As central banks set short term rates, all of the term structures shift: quantitative analysis of the term structure movement shows that this parallel shift explains most of the movements in the term structure. The other two main movements are slope and curvature7. Exhibit 5 shows these three movements of the term structure of interest rates. Unfortunately in a situation where short term rates have limited freedom of movements, the situation might prove radically different8. Exhibit 6 shows the sensitivity to interest rate movements of different points of the term structure, namely of two, five and ten years for the US and Europe. This clearly illustrates a change in behaviour catalysed by the introduction of very low interest rates. Again this can have a large impact on a systematic investment strategy exploiting bond trends, as a short-term contract might react in a way not accounted for by any historical time series.
Exhibit 5: The main moves of the interest rate term structure: parallel shift, slope, and curvatures
Interestingly, the two effects can interact. In fact, as the volatility of short-term contracts decreases, in an inverse volatility weighted allocation methodology, their allocation increases. At the same time, as discussed, their behaviour changes radically. The strategy exposure might therefore be altered and attention should be paid to understanding whether it is still able to fulfil its role of efficiently harvesting the momentum risk premium in the bond market.
Exhibit 6: Sensitivity to interest rate movements of different points of the term structure, namely of two, five, and ten years for US and Europe
As this discussion shows, quantitative investing and in particular harvesting risk premia is not only a matter of constructing an ‘in sample’ generic strategy and letting it run: past data tells only a part of the story. A backward and forward looking approach, with a sound understanding of the harvesting algorithm and all its implications is, in our opinion, of paramount importance. As always in finance the devil lies in the details and neglecting those leads to poor models and sub-par performance.