Machine learning - Rise of the machines?
Simon Ho, Senior Analyst - Alternatives, MLC
Machine learning is amongst one of the most promising and talked about technologies in recent years.
With the exponential growth of data being created and capacities to process those data, machine learning has been a game changer to fields like retail marketing, driverless cars and manufacturing. Given its success in these areas it seems an ideal fit from an investment point of view especially for systematic or quantitative investing. But has it had the same impact? And, what does it mean for investing going forward?
What exactly is machine learning?
Academically speaking, machine learning (ML) is centred on the idea of algorithmicmodels having the ability to 'self-learn' without being explicitly programmed on how to do so. 'Self-learning' refers to the algorithm’s ability to improve predictive accuracy over time as we feed it more realised data. Practically speaking, what sets ML models apart from classical statistics models is its ability to identify more nuanced and complex relationships (often in the form of non-linear relationships) between what we want to predict (the dependent variables) and the predictors. This ability to identify 'second order effects' has become increasingly important for systematic managers, who focus on identifying market patterns in a methodical fashion, as more and more funds compete in the same space for the same alpha. Every bit of advantage helps.
Another major development contributing to the rising popularity of ML is the proliferation of new data, which stems from an increasing digitisation of our daily lives. Many investment nuggets can be unearthed within the new data available to us, but we need to know ‘how’ to look. The data is often large and unstructured in nature (think each pixel in a satellite image, unstructured transcript of an earnings call etc), and classical statistical techniques just aren’t built to deal with these kinds of datasets. This is where ML comes in. ML models are usually algorithmic in nature, and therefore make few assumptions about what the data should look like compared to classical models. As a result, ML models are much more powerful in identifying insights from unstructured datasets, which are increasingly utilised by systematic managers.
Financial markets are a different beast
It sounds like ML is a superior analytical tool more suited to contemporary times. However in reality, it has not quite come to dominate investment decisions. This is because of two ‘distinct’ features of financial markets; non-stationarity, and low signal-to-noise ratio of market data.
Unlike natural science (eg physics), there are no strong underlying laws that govern financial markets. It evolves and adapts to its participants’ actions very quickly. This becomes a problem when we try and make predictions based on past relationships, which is essentially what ML does. As we all know, market dynamics can change quickly (we are currently in a period of unprecedented elevated risks), rendering any past relationships useless in terms of predictive power into the future. This is the phenomenon of non-stationarity in the financial market’s underlying mechanism.
Secondly, financial data has extremely low signal-to-noise ratio. The volume of market and economic information is increasing exponentially, but the amount of useful information isn’t. Large amounts of data doesn’t mean large amounts of useful information. Much of it is noise. Signal is truth that should be taken seriously, but noise distracts us from the truth and often leads investors down rabbit holes.
Furthermore, machine learning techniques give investors more flexibility to fit past data closely; this is both a blessing and a curse. It can easily ‘over-fit’ our data if we are not expertly attuned to the workings and limitations of the ML technique being applied. Overfitting happens when a model relies on past data too much, negatively impacting its ability to analyse and successfully use new data.
Combined with the issue of non-stationarity and low signal-to-noise ratio of the underlying data, it is not difficult to see how easy ML can exacerbate over-fitting risk and rely too heavily on past data which has little relevance to the future.
In many areas of investment decision-making, the susceptibility to over-fitting outweighs the many benefits that ML may bring, which has been the main reason why ML hasn’t taken over most aspects of systematic investing yet. For example, ML is not often used as the overarching alpha model, because asset returns are known to be notoriously non-stationary. Instead, economic rationales are baked into the overarching models, in order to give the model a bit more structure. A relatively timeless structure that serves to combat over-fitting.
Where in an investment process does it fit?
The trick is to be selective in the problems we apply ML to. Systematic managers have applied it to certain subsets of their investment process where the underlying problem is more of a stationary nature, and large amounts of quality data is available for the ML algorithm to learn from. This ensures predictive accuracy is improved whilst over-fitting is mitigated. Below are a few examples of popular areas where systematic managers leverage ML techniques:
- Help better predict earnings-based forecasts – Whilst stock returns are quite non-stationary to forecast directly, earnings-based metrics (eg earnings yield) on the other hand are relatively stationary. Therefore, we can often use ML models to predict earnings growth, and in turn, use earnings growth to make inferences about stock returns.
- Feature generation - ML algorithms are often used to extract certain ‘features’ in an alternative dataset. These features may then be used to forecast earnings, which in turn act as an estimator to asset prices. For example, ML is often applied to satellite images of department store parking lots; the algorithm’s main goal is to determine what constitutes a car, thereby generating number of cars as a feature, which should have a positive correlation to sales forecasts. Another use is the analysis of geo-location data to monitor foot traffic into retail stores, which is in turn used to predict sales ahead of its company earnings announcement.
- Sentiment analysis – classification techniques such as neural networks can be used to help classify public sentiment of a company’s product using social media data feeds, for example. Or the sentiment of a company’s earnings announcement.
- Execution systems – ML algorithms are often utilised to help managers execute their trades more effectively, thereby reducing the portfolio’s trading cost. ML is particularly suited as execution models because of the vast amount of data generated via ticker feeds that have a high signal-to-noise ratio.
- Improve current model robustness – ML models are generally quite flexible, therefore it is fairly susceptible to over-fitting risk if not implemented carefully. For this reason, ML researchers have made important advancements by discovering a plethora of techniques specifically designed to combat over-fitting risk. Examples in this category are cross validation, ensemble models etc. These auxiliary techniques are also borrowed by systematic managers to make traditional alpha models more robust.
Revolutionary? Maybe, but not yet
Machine learning technology has made significant advancements in sectors such as manufacturing, autonomous navigation, and consumer behaviour prediction, where underlying processes are relatively stationary. However, the investment world is a different beast and the application of ML techniques alone is not easy due to the presence of non-stationarity of asset prices and low signal-to-noise ratio of financial data.
Instead, most professional systematic investors complement their traditional investment models with machine learning algorithms to areas where predictive accuracy is improved while over-fitting risk is contained.
Investors should be sceptical if a systematic investor (barring high frequency traders - they’re a different kettle of fish) claimed that they heavily utilise machine learning in their investment process.
In short, machine learning is evolutionary as opposed to revolutionary for systematic investing (at least, not yet). Machine learning is another set of tools in a quantitative investor’s toolkit that he/she should understand well, in order to help navigate this increasingly competitive and unprecedented investment landscape.
This communication is provided by MLC Investments Limited (ABN 30 002 641 661, AFSL 230705) (MLC), a member of the group of companies comprised National Australia Bank Limited , its related companies, associated entities and any officer, employee, agent, adviser or contractor (‘NAB Group’). An investment with MLCI does not represent a deposit or liability of, and is not guaranteed by, the NAB Group. The information in the communication is of a general nature only and is not financial product advice. The communication is not intended to offer products or services provided by any member of the NAB Group. Opinions constitute our judgement at the time of issue and are subject to change. Neither MLCI nor any member of the NAB Group, nor their employees or directors give any warranty of accuracy or reliability, nor accept any responsibility for errors or omissions in this communication.
Share this article: