Using the TIMEMIXER approach for volatility forecasting

An application of time mixing is in the area of volatility forecasting. Research has extended the work on GARCH to an extreme, but there may be other techniques in time series forecasting that can be applied to financial time series that may be very useful. A recent paper focused on TimeMixers which employs different time scales as a method to improve forecasts. See "Volatility Forecasting in Global Financial Markets Using TimeMixer".

The idea behind TimeMixers is straight-forward. There is imbedded in any times series relationship with different timeframes that can exploited. There can be long-term seasonality. There can be cycles or trends that are longer than a few days that will not be captured with daily data. Classic time series in ARMA models can handle seasonality and can identify autocorrelation at different lengths, but a more explicit breakdown of data may improve forecasts. 

I like the technique used and the author applied it to a broad set of markets, but I was disappointed that there was no testing against other types of models for volatility. This process looks interesting but it is not clear it is any better than what we already have. The MAE, MSE, and RMSE all are low especially for short-term forecasts but the quality of technique has to be with the results, the easy of understanding and the ease of implementation. This paper does not make that case for TimeMixer ML. 

That sinking feeling in the CMBS market

 

Sometimes you get hit with a wave, but often you drown because the water level just gets too high and you don't have the energy to find the shore, or worse, there is no shore. There is a problem in the CMBS problem with some very good properties underwater with the sinking still occurring. We can start with the issue of running a CMBS on a single property. There is no diversification. Investors may expect that lower rated tranches will have more credit risk, but in these cases, we are seeing risk with the A tranche. Ouch. This is what is supposed to be avoiding with structuring. A combination of poor engineering with a bad environment spells disaster for investors. 

Can there be money to be made? Yes, however, it is not from dropping these bonds in a portfolio and cutting some coupons. This is risky bond trading. 

TimeMixers as a new time series tool

 

Many trend-followers are time mixers. What has been old for some is now something new. A new technique in machine learning time series forecasting that is being given more attention is a concept called time mixing. The idea of time mixing is simple. Decompose any time series into a set of different time scales. For example, there can be short-term behavior, longer-term, seasonal and very long-term behavior in prices. This breakdown is certainly true for almost all commodities. If we can take the multi-scale data, we can break the problem into two parts or blocks:  the past-decomposable-Mixing block PDM where the information at different scales is learned by the model. From the PDM, the model is sent to the Future-Multipredictor-MIxing (FMM) block which will combine the scaled information into a forecast. This type of model may not do well for very short-term predictions, but longer-term forecasts can be improved by this technique. Given this is a ML technique, the value is found in improved forecasting. This is not a structural model of how the market behaves.

What do many trend-followers do? They will often use trends for different scales or look-backs to blend into a single signal. The weighting of these signals can be a very simple, like an equal-weighted scale, or it can be optimized based on preferences or on past forecasting success. This is not the sample as a formal time mixer model, but the idea behind this approach is the same. Use different time scales to capture the different behaviors in asset time series. 

Causal structures for equity risk factors evolve over time

 

There has been a strong increase in the study of causal inference in finance. Given the increase in the factor zoo and p-hacking, there is a need to go back to basics and think about causality and what it takes to find causal relationships. 

The paper, "The Evolving Causal Structure of Equity Risk Factors" focuses on using causal structure learning methods to analyze 11 risk factors in the US equity markets. The relationships have become more sparse with the market beta having a strong influence on other risk factors. Much of this research work starts with structural vector autoregressive models which look at the impact past returns on current returns to measure the impact of cross-factor causation. SVAR models can be restructured as directed acyclic graphs (DAG) that can provide useful insights on causality. Inferring causality from these insights is the objective of causal structure learning. 

The number of edges for both correlation and causal networks have fallen through time. The markets are more sparse and more focused on instantaneous relationships. The authors also find that the number of edges in the network is negatively related to VIX shocks; however, the number of edges in the causal network will increase with a volatility shock. The volatility shock is a large increase (above the 95th percentile) in VIX between the close and open values standardized. 

This work is just a start. What is more important is finding casual relationships that can be exploited for profit.

The evolution of the hedge fund industry - The development of multi-strats

 

The hedge fund industry has gone through significant changes over the last 20 years with a recent paper providing a good description of how hedge fund structures have adapted especially since the GFC. See from Clear Alpha Technologies "The Evolution of Alpha: Exploring the past, present, and future of investing in alpha". I don't like the title because this is not about alpha creation but how an industry has adapted to changing tastes by the market to deliver what is expected to be a better alternative investment product. The alpha creation from the bundling of strategies is different from how returns are delivered to investors which is the topic of this paper.

The hedge fund industry started with the development of specialized money management funds which usually had a focused objective. From traditional long-only managers came the broader focus of long/short managers. Unfortunately, there are many ways to generate return and alpha, so investor demand bundled products in the form of fund of funds. Unfortunately, many fund of funds did not delivery on the expected promise of higher returns and controlled risk. 

Fees were higher relative to what was delivered so the market looked form other alternatives. One approach was to bring the investing in-house as market knowledge of hedge funds increased. The suppliers of returns developed in a different direction through creating multi-strat programs under a single structure. This was extremely successful, but there were limited in the ability of these multi-strats to gather the investment talent relative to demand. A new development was the multi-strategy, multi-manager fund which blended internal and external managers in a single structure. In both cases, the multi-strategy approach allowed for better risk oversight and more effective management of capital allocation. 

The other development from the multi-strategy structure was the pass-through of costs. This pass-through model protected the hedge fund business and had the ability to provide a better cost structure to investors if the costs could be controlled. Investors were now investing in hedge fund businesses not giving money to be managers. There is a nuanced difference but has a significant impact on both the suppliers and demanders of alpha. The suppliers can invest in technology and risk management and reduce fluctuations in revenue. Investors, the demanders of hedge fund services, can form strategic partnerships but may have to pay a premium for the services of these multi-strat managers. 

Hedge fund management is a very dynamic industry which will evolve as alpha changes, and the management structure will impact alpha. 

Dynamic Factor Allocation with Regime Switchng

When should you change the allocation of your factor exposure? This is a critical question and ultimately it is a question of market timing. But how do you effectively market time? This timing can be done through looking for regime switches. If we can identify if we are in a bull or bear regime, we should be able to adjust the allocation across factor allocations. Of course, if we could be perfect timers, we would not think about factor exposures, we would just go long the risky asset in a bull market regime and go short in a bear market. Alternatively, we focus on building a portfolio because we know that we will not be perfect in our timing ability.

In the new paper, "Dynamic Factor Allocation Leveraging Regime-Switching Signals", the authors look at seven long-only indices which include the market portfolio and value, size, momentum, quality, low volatility and growth and employ a Black-Litterman model with regime switching. They show that the dynamic portfolio will do better than an equal-weighted portfolio when the allocation decisions are based on a sparse jump (SJM) model. The SJM uses risk and return measures to identify the regimes. It is a form of cluster analysis to find bull and bear behavior across the set of factors analyzed. This regime model is the novel addition to this paper. A regime model is applied to each factor, so each has identified their own bull and bear period.

If you can identify regimes correctly, you can market time. That should seem obvious, and it is. There are many ways to form regimes, so the key job of the PM is to find the right way to identify these regimes.

Note that different regime modeling techniques will generate different regimes for the same data. The regimes for the size factors are very different.

Whether this can be implemented in world of transaction costs, is questionable, but the key concept of using regimes to help build portfolios is very useful.