Machine learning - what it is and is not for investors

I have been seeing more managers use machine learning tools to help with the investment process. This is an important advancement and will be very useful in helping generate better return generating engines. However, there are may investors who do not known what machine learning is or what it should be able to do. Simply put, machine learning is having an algorithm learn without explicit programming, a process of improvement with experience from new data. It is the training of a model for data that can be generalized for decisions against some performance measure. 

I have listed some key ideas of what machine learning is and is not for an investor who is just being exposed to the concept.

Machine learning is not artificial intelligence.

Many think that machine learning is a sub-discipline within AI; however, there has been a large divergence between machine learning which is grounded in data and statistics and AI which is focused on logical systems. AI, for many, never realized its potential in the investment area. We do not believe that the same failure will exist in the machine learning field given the different goals and objectives.

Machine learning is not data mining.

Data mining focused on finding relationships in large data sets that were not able to be easily discovered or are not readily apparent. Machine learning potentially uses large data sets to train systems to make predictions. The elements of machine learning include: data mining, statistical inference, and prediction.

Machine learning is not a black box.

Many AI models have relied on neural networks which can seem like a black box which cannot be understood by the manager or investor. There is talk of a “hidden” layer which adds mystery. Machine learning can be closely directed or supervised in order to apply what has been learned to new data, or it can be unsupervised in order to draw inferences from data.  

Machine learning is not new.

Machine learning was first developed decades ago with many of the statistical techniques used coming from older approaches toward inference and prediction. The more recent development is that cheap computing and large data sets has allowed for more and quicker learning.

Machine learning is not testing endless alternatives.

Machine learning does not mean that a computer and manager is going on an endless hunting expedition for an over-optimized result. In fact, the value of machine learning is the ability to blend and weigh many alternatives which could not be done in the past. 

Machine learning requires statistical/programming/market skill.

Fundamentally, the manager who uses machine learning combines strong statistical foundation with programming skills to look at a  wide set of alternative which can be weighted or excluded. Nevertheless, there still needs to be a sense of market to help learning and understanding.

Machine learning will find relationships you did not expect.

Machine learning is not just data mining which is explicitly looking for relationships within a large data set; nevertheless, machine learning may be able to weigh different model alternatives and find combinations which would not be thought of through the normal process of simple hypothesis testing. There is learning without explicit programming.

Machine learning does require a lot of computing power.

The machine learning by its very nature will look at many different combinations of data and learn how to update models when new information is introduced and patterns found. The process of stepping forward and determining the impact of new information is computing intensive.

Machine learning can be either supervised or unsupervised.

While machine learning can be unsupervised and free form when looking at model alternatives, it can also supervised whereby the learning is very specific and directed to apply what has been learned from new data.

Machine learning will make you a smarter investor.

When machines learn, the modeler learns. One of the key advancements of using machine learning is the ability to find dynamic relationship in data which are not immediately obvious and make predictions.

Structural checklist for investment headwinds

Systematic models are very useful and important for disciplined investing, but the

percentage of the variation explained by most models in asset markets is relatively low especially over short horizons. The question for most managers, even quants, is determining how to deal with this percentage that is unexplained. There is no simple solution but applying a structural headwinds/tailwinds checklist may provide a good first pass for addressing the problem.

A structural headwinds or tailwinds checklist groups or categorizes issues that may provide a tilt to returns. These tilts on expected returns may lead investors to make a tilt or base adjustment to asset class allocations. Instead of starting with a base allocation of zero, there could be a negative or positive base allocation.

There could be more categories to this checklist, but these eight may get any discussion started. Some of these issues can be quantified, but we believe they often cannot explain short-term variation except if there is a shock. These factors can have an impact on returns through impacting the risk premiums in markets but only over long horizons.

Structural headwinds - A checklist

• Demographics - 
• The era of aging is upon us and it has an impact on capital flows and savings rates; just ask Japan or Europe. Demographics may include issues like the flow between rural and urban areas in China. It will drive return patterns even though it will not affect short-term volatility.
• Government -
• The type of government that is in place will affect investment options. Government impacts could include gridlock.  Venezuela is a perfect example for where government matters. The same can be said for Argentina or Russia.
• Regulation - 
• The Dodd-Frank regulation has an impact on returns but it is hard to model the impact directly. All of these rules on banks impact credit and the financial sectors. More regulation will have a greater impact on small cap stocks.
• Elections -
• BREXIT will impact all of the globe. The US presidential election may have a profound change on trade and global relations. The election dates are well known and may outweigh other model factors.
• Geopolitical risks (war - terrorism)-
• These risks usually lead to shock effects, but the changing probabilities of geopolitical risks will impact all returns; nevertheless, it is hard to include in any model. 
• Global trade -
• This factor refers to the overall integration of capital, labor, and good around the globe and not the trade balance of any one country.Globalization will impact capital market integration will effect the correlation across markets. 
• Climate - weather-
• While there has been much talk about climate change, the impact on investment returns is less clear-cut. Obviously, there are weather events that impact returns, but these can be diversified. 
• Technology - 
• Technology can provide a boost or a drag on specific industries. In the global macro arena, the impact of technology is less clear but the impact of technology on finance is real and does affect liquidity which is being priced in the markets.

Even if these factors are not explicitly modeled, they will affect allocation decisions and should at the least be catalogued and discussed.

Managed futures - something going on with performance

It is always worth monitoring the long view with trends. Beyond the last quarter or year, there are return patterns that show longer-term changes in style performance. Take a look at managed futures. There was a long period for which it was out of style, but that has changed markedly. (Saying a strategy is out of style is the nice way of saying it has performed poorly.  In style, means it is doing well and investors are chasing performance.) 

We fitted a long-term linear trend through the performance curve of the SocGen managed futures index. The times of strong or poor performance from trend are obvious by looking at periods above or below the trend line. A fitted polynomial through the data can find periods when the slope of returns was rising or falling and provide a smoother view of performance.

There are clearly long periods of strong and poor performance. There is an uptrend in performance through the Financial Crisis, but the slope turned down during the  post-crisis or QE period. Nevertheless, the second half of 2014 marked a new period of performance. This new era is completely different than past periods. (We looked at log value and see the same pattern.)

Of course, we fitted the entire period which would not be available to investors, but the story holds. If an investor looked at holding a managed futures basket when the non-linear curve slope is rising for the index and selling during periods of declining slope, you would have have positive style tailwinds. This is not a model,  but may be a good rule of thumb. 

Stop the mistakes - play the odds from a process

When you repeat a mistake, it is not a mistake: it is a decision.
- Paulo Coelho

So what is an investment mistake? If the odds are 60/40 in your favor and you lose, is that a mistake or are you just unlucky? In this case, a bad outcome is just bad luck. If you calculate the odds to be 60/40 against you and you have a gain is that a lack of skill or just being lucky. These are important issues to consider if you want to measure or show skill. 

A mistake is not being rational or being inconsistent with an articulated decision process. If there is a process and the decision-maker deviates from the process, it is a mistake regardless of your luck. If you don't have a well articulated process, you will not be able to say that you made a mistake. A more fluid process allows for ambiguity and thus makes it hard to admit a mistake even during bad performance periods.

Given there is a process, the real issue is whether the decision-making model is wrong. If you are rationally following the model, the focus is on whether the model of market behavior is incorrect. If you consistently follow a disciplined process but if the model's predictive power is flawed, there is a mistake. Unfortunately, most models of asset prices can only explain a small portion of the variation in prices. Hence, mistakes can be repeated and turn into bad decisions.

Being systematic is just a start. It is the predictive process for which discipline is applied that really matters. Even here a trader faces a severe problem because mistakes are generally more frequent than success. A higher success ratio does not mean high returns. Success could be less than 50% yet produce higher returns. You can make repeated mistakes and still profit through controlling risk. A flawed model can still be effective if the trader cuts his losses. 

Mistake happen but decisions matter.