Do you have any queries about NMRQL Direct? If you don’t find the answers below, you’re welcome to contact us.
What is your philosophy on the markets?
Asset prices are determined by supply and demand. In financial markets supply and demand are determined by market participants’ views which are more often than not influenced by data including: the news, fundamentals, technicals, and economics. Financial markets, therefore, act as a capital-weighted consensus mechanism; they blend all of the market participants views together and what we get out is asset prices.
The first “problem” with this mechanism is that people aren’t rational; the way we process data is unreliable and changes over time. The second “problem” with this mechanism is that capital tends to accumulate which causes particular views to dominate the market. This is why we see bull markets and bear markets. Bear markets occur when pessimistic views are dominant whilst bull markets occur when optimistic views are dominant.
At NMRQL we train hundreds of machine learning models on different sets of data including fundamental data, technicals, and economics. These machine learning models generate predictions and are then put into a voting mechanism which blends them together.
In some sense what we are trying to do at NMRQL is emulate what the capital-weighted market consensus of the market will be for each asset. Machine learning allows us to do this in a cost-effective, adaptive, unbiased, scalable, and historically-testable way. By that we mean:
- Our algorithms don’t get paid high salaries and bonuses.
- Our algorithms’ views change over time.
- Our algorithms do not suffer from cognitive biases.
- Our algorithms can process any form of data a human can.
- Our algorithms can be simulated historically to see how they would have behaved.
What is machine learning and how does it work?
Machine learning is the study of algorithms which are able to learn how to solve a problem given data.
Most machine learning starts with inputs, desired outputs, and an algorithm whose behaviour can be tuned. At first, the algorithm processes the inputs randomly to produce outputs. When the error between the outputs from the algorithm and the desired outputs is high, the algorithm is penalized more, when it is low, the algorithm is penalized less. Over time, the algorithm tunes its parameters until it can produce the desired output.
At NMRQL Research the inputs into the machine learning models may include the news, fundamentals, technicals, and economics and the outputs are future returns. For more information about how machine learning is done at NMRQL see our machine learning page.
How do your machine learning models handle market irrationality?
Each of our machine learning algorithms are continuously learning and being scored. When the algorithm score starts to deteriorate this is an indication that what is has learnt is no longer valid. In this situation, the importance of the algorithm in the “organization” of models is reduced. When all of the algorithms’ scores deteriorate, that is a signal for us to move assets from riskier investments to less risky investments because the market is irrational.
How does this relate to other forms of investing?
Our investment philosophy is consistent with traditional forms of investing such as value investing, passive investing, trend following, and quantitative investing. Markets are capital-weighted consensus mechanisms. Because capital tends to accumulate over time to certain investment philosophies, markets change. Therefore, we believe that most investment philosophies fluctuate between periods of being “right” and periods of being “wrong”. That is to say, periods of being the consensus and periods of being contrarian.
How does this differ to quantitative investing?
Traditional quantitative investment strategies typically start off with an assumption about how a market or particular asset works. Machine learning starts with no prior assumptions. Everything the algorithm knows is learnt from the data without interference from human dogmas.
For example, a quantitative investment strategy might start off by assuming that markets exhibit something called momentum and go from there. Our algorithms do not work this way, they will learn about momentum if, and only if, it objectively exists in the raw data.
Aren’t markets supposed to be efficient?
Market efficiency is typically defined in terms of two variables: a set of information and a set of models for processing that information. At NMRQL Research we believe informationally efficient markets are highly improbable because:
- Very few models can process any type of information e.g. video or sound,
- Whilst humans can process any type of information, this processing is typically biased because we suffer from a number of cognitive biases, and
- Markets are unlikely to be efficient with respect to current-generation machine learning models because very few market participants are using them.
Furthermore, informationally efficient markets were proven to be impossible according to Grossman and Stiglitz in 1980. In addition to this a number of statistically and economically significant “anomalies” have been observed consistently in global financial markets.
Is this high frequency trading?
No. Our algorithms generate longer-term predictions and trade slowly.
Isn’t machine learning a black box?
Machine learning is complicated, but that does not mean it is a black box. At NMRQL we understand the maths, but more importantly we have developed a number of tools which allow us to visualize and understand what our models are doing.
Is machine learning used abroad for investment management?
Offshore, there are a few hedge funds that have led the way in the use of machine learning algorithms in their funds for trading, data scrutiny or pattern recognition applications. There is also a growing number of larger traditional managers offshore that are starting to use machine learning algorithms. We believe that adoption in South Africa has been slower than offshore for the following reasons:
- Many breakthroughs in machine learning have only occurred in the past 3 – 5 years at the likes of Stanford, Google, and Facebook.
- The amount of data available to learn from is growing exponentially. Generally speaking the more data an algorithm sees, the better it gets. Historically there was probably too little digital data available for machine learning algorithms to be effective.
- The skillset required to design and implement machine learning algorithms – especially in dynamic and uncertain environments like financial markets – is extremely broad and the talent pool is extremely shallow.
That said, we are very confident that machine learning is going to be become a more prominent investment strategy in South Africa going forward.
What is the nature of the relationship between NMRQL Research and Sanlam?
As a small boutique asset manager, we do not have the resources to manage fund administration. We therefore rely on a collective investments platform to do this on our behalf and co-brand (or white label) our fund on the Sanlam Collective Investments Platform. From a legal point of view, the fund is administered and managed by Sanlam Collective Investments, who then outsources the management and decision making on all investments to NMRQL Research. This provides the investor with nimble and innovative decision making using new technology, while still maintaining the peace of mind provided by a larger entity with a large balance sheet and solid compliance, risk and reporting capabilities.
Why did your hedge fund underperform in the first year?
No new technology is ever successfully implemented straight away. Thomas Edison failed to invent the lightbulb more than 1,000 times. The Wright Brother’s failed to fly numerous times. SpaceX’s first three rockets blew up. We are no exception to that rule.
Our hedge fund started trading with our own capital in August 2015. In the two years since then we have made mistakes, but much more importantly, we have learnt extremely valuable lessons and are now confident in our algorithms and risk management processes.
How do you manage financial risk?
Our algorithms reduce market risk by maintaining a strict asset allocation, not investing too much in any one stock, and studying correlations to avoid being too concentrated in similar stocks. Our algorithms reduce liquidity risk by only investing in the top 50 assets listed on the JSE by market capitalization. Our algorithms also consider transaction costs and slippage.
Another form of financial risk is model risk – the risk that your model is wrong. At NMRQL we manage this risk first and foremost by adopting strict software engineering principles and secondly by diversifying over the outputs from 100’s of uncorrelated models.
What is the NMRQL SCI Balanced Fund’s objective?
The fund aims to achieve steady long-term growth of capital and income by investing in a diversified portfolio of domestic and international assets, where the asset allocation and stock selection is systematically managed using machine learning algorithms.
What are your asset allocation allowances?
For information on the NMRQL Multistrategy Prescient QI Hedge Fund please contact us directly. The minimum investment is R 5 million.
The NMRQL Sanlam Collective Investments Balanced Fund is a regulation 28 compliant fund meaning that it is suitable for pension funds.
As such we are allowed the following asset allocations:
- A maximum of 75% in equities (stocks)
- Of the 75% a maximum of 25% in offshore equities
- A maximum of 10% per stock with a market cap greater than R20 billion
- A maximum of 100% in a listed cash or cash equivalent
- A maximum of 25% in any one cash or cash equivalent instrument
- A maximum of 100% in any one SA government bond
- A maximum of 10% in commodities
- A maximum of 5% in any one commodity
- A maximum of 25% in listed properties
- A maximum of 80% in SA Collective Investment or ETF
The total exposure of the fund to investment assets may not exceed 100%. The unit trust fund may also not implement any naked shorts in the portfolio.
How do you manage technology risk?
We follow a very strict software engineering methodology and all of our code is version controlled, unit tested, regression tested, and simulated historically before being released into the wild. That said, all code has bugs so the fund managers look at trades generated by the algorithms before implementing them. Our algorithms don’t have direct market access.
How do you simulate your machine learning models?
We have developed a cutting edge collective intelligence framework at NMRQL called recipes. Recipes is designed to automatically remove data and algorithmic biases such as survivorship bias, look-ahead bias, sampling bias, and curve fitting behaviour.
Recipes allows us to simulate how our machine learning algorithms would have behaved in a walk-forward fashion. Under this approach, the algorithms are evaluated at each point in time historically using only the information that would have been available to them at that point. At each point in time they learn from the data and optimize their own internal parameters. This approach mitigates the risk of curve fitting and data mining.
Can I see a backtest / historical results?
No, we are not allowed to show this according to South African regulations.
How did you choose your benchmarks for the funds?
The benchmark for the NMRQL Sanlam Collective Investments Balanced Fund is the ASISA Multi Asset High Equity peer group average. At NMRQL we believe that when everybody does well (perhaps due to favourable market conditions) we should not be charging high fees. By using the peer group average as our benchmark NMRQL is only rewarded if we beat the average manager, when everybody is doing well this is harder to achieve.
The benchmark for the NMRQL Multistrategy Prescient QI Hedge Fund is cash (STEFI) because the objective of the fund is to provide a source of good risk-adjusted-returns to our investors which is uncorrelated with the stock market as a whole.