Technology improves exponentially. In 1945 a project which cost $6,740,000 in today’s terms built the world’s first computer, the Electronic Numerical Integrator and Computer (ENIAC). The ENIAC housed 17,468 vacuum tubes and took up an entire room. In 1971 Intel released the world’s first single-core microprocessor, the Intel 4004, which was more than 10x faster than the ENIAC and could balance on the edge of your fingertips. The Intel 4004 had a processing speed of 0.00074 GHz and fitted 2,300 transistors. In 2014 the iPhone 6 was launched with the dual core A8 processor which had a max clock rate of 1.1GHz, almost 1,500x faster than the Intel 4004 and which – according to some wild calculations – would take 10,000,000 ENIAC’s to emulate. I won’t even attempt to compare the ENIAC, Intel 4004, or even the Apple A8 to Nvidia’s new GeForce GTX 1080 Ti which packs an incredible 12,000,000,000 transistors!

Exponential improvement at lower and lower costs has made technology ubiquitous. It is in interstellar space, orbiting the earth, at the bottom of the ocean, flying our planes, driving our cars, in our fridges, TV’s, pockets, and even on our wrists.

Now consider the fact that many of those devices are connected to the internet and are producing digital data. By some estimates we will create more data this year than was produced in the last 5,000 years! Furthermore, an incredible 70 – 80% of that data is unstructured meaning that it exists as websites, social media, voice recording, images, and video. Traditional investment techniques, which rely on humans processing all the relevant data and even systematic investment techniques, which rely on traditional programming techniques and database technologies, will inevitably fail to exploit unstructured data because of its sheer size and lack of inherent shape or form.

The good news is that there is one technology which stands a chance of making sense of our exponential times and it is being developed by the most forward-thinking and innovative companies of our age including, but not limited to: Google, Facebook, Amazon, Netflix, and Tesla. This technology is called machine learning and it is the heart of our business at NMRQL Research. Our investment philosophy at NMRQL Research is bold and simple: continuously do machine learning at an unprecedented scale on exponentially increasing sets of structured and unstructured data.

Allow me to explain why this counts as an investment philosophy.

Finance 101: What is in a price?

The price of a security in a market is determined by supply of and demand for that security. When enough market participants want to buy the security, its price goes up. When enough market participants want to sell the security, its price does down.

Almost all of the time, market participants buy and sell when they believe prices will go up or down in the future. In other words, most of the time supply of, and demand for a security are determined by market participants views of the future.

Most market participants arrive at their views of the future buy analyzing some set of information using some set of models. Here are some over-simplified examples:

Value investors analyze company information using valuation models such as dividend discount models. Technical traders analyze historical price action using technical indicators such as bollinger bands. Macro investors analyze information about large-scale economies using macroeconomic models. Large insurance companies analyze demographic data using actuarial models. Multifactor investors analyze swathes of cross-sectional data using filters and ranking methods …

These belief systems are called investment philosophies. Interestingly, a growing number of market participants maintain no view of any individual security’s future and prefer to buy the index. This involves no view formulation or information processing.

The market allows participants to buy and sell any given security based on their views. In this process the market blends those market participants views together and what we end up with is a price which – to some degree – reflects both the set of information and the set of models used by participants to formulate their views.

Ultimately, the degree to which any set of models or set of information is already reflected in the price of securities for a given market, depends on how much capital is allocated to strategies using those sets of models and sets of information.

The Great Debate: Are Markets Efficient?

Yes, maybe, and no. Consider this definition proposed by Timmermann and Granger in their paper: Efficient Market Hypothesis and Forecasting:

A market is efficient with respect to the information set, Ωt, search technologies, St, and forecasting models, Mt, if it is impossible to make economic profits by trading on the basis of signals produced from a forecasting model in Mt defined over predictor variables in the information set Ωt and selected using a search technology in St -Timmermann and Granger

Let’s break that down a bit. Market efficiency is defined with respect to four variables and one concept. The four variables are forecasting models, information sets, search technologies, and time. The concept is economic profits i.e. profits after all costs.

What constitutes forecasting models?

Most investment decisions are still made using the human brain. Don’t get me wrong, the human brain is a wonder. Just consider what most of us can do and take for granted. We can see, hear, touch, walk, run, jump, climb, drive, think, speak, sing, read, write, dance, compose, and inspire. Machine learning algorithms have yet to truly master any of these, let alone master them concurrently (and take it for granted).

That said, our intelligence is better suited to creative endeavours than investing in the markets, where remaining calm and rational is benefitial. When we look at the clouds we see dogs and cats, whereas algorithms see random configurations of particles. When we look at stocks we see “patterns”. Algorithms see only what the data unequivocally supports. We also suffer from biases and get tired after a long day.

Investment decisions could also be made systematically. Whilst this reduces bias and increases the capacity to process structured data, most systematic investment strategies are left wanting when it comes to raw unstructured data. These systems are also unable to learn and adapt as the market changes and are therefore left clinging to dogmatic assertions, long after strategies have outlived their usefullness.

We constitute forecasting models as cutting-edge machine learning algorithms. We have done this because algorithms are inherently unbiased, machine learning algorithms can adapt to change, and they can learn from any form of data.

At NMRQL Research just one of the models we have coded up and use are Long Short-Term Memory (LSTM) Neural Networks. These are the same class of Neural Networks powering most state-of-the-art Neural Machine Translation systems.

What constitutes information?

The investment philosophy a firm chooses also dictates what they constitute as information. Value investors look at cashflows and fundamentals. Multifactor investors look at premia like size, value, growth, and momentum. Technical analysts look at transformations of price and volume time series. But why stop there?

Using natural language processing it’s possible to “read” millions of tweets, news articles, and announcements a day to extract investor sentiment or consumer confidence. Using image recognition it’s possible to “watch” the number of ships coming in and out of the port of Cape Town to better estimate imports and exports.

We constitute structured and unstructured data as information. We have done this because there is information in structured data which traditional forecasting models miss and because unstructured data contains useful, not-yet-known information.

Using machine learning algorithms it is now possible to understand unstructured data such as naturaly language. We can identify what it is about and whether or not it is positive.

What constitutes search technologies?

Search technologies are what we use to search through the myriad of possible combinations of information and models. The most prolific search technology being used in the investment industry today is still the organization: thousands of analysts trying new things (usually in Excel) to get a “collective gut feeling” for what works and what doesn’t work. Needless to say this search technology is not very scientific.

Now consider that the capacity of your smart phone is millions of times more powerful than the computers which were used to land humans on the moon.

Computing has never been more powerful, less expensive, and more available than it is today. Programming techniques allow us to automate boring and repetitive tasks such as searching through possible combinations of information and models. Added to this, cloud computing has made it possible for anybody anywhere in the world to spin up massive CPU and GPU clusters at a fraction of the cost of buying hardware.

We constitute the latest and most available computing services as search technologies. We have done this because it allows us to search through more combinations of models and information in a more scientific way and at a much lower cost. Leveraging programming and computing also means that the humans who work at NMRQL Research will have fewer repetitive jobs and hopefully more stimulating careers.

This simple demo of Cloudways CDN shows just how easy it really is to spin up massive CPU and GPU clusters on a variety of different cloud computing platforms.

What is the significance of time?

Timmermann and Granger are smart. They realized that because information sets, forecasting models, and search technologies change over time (and when speaking about technology they change exponentially over time) market efficiency must also change over time. In other words, because most investors are slow to adopt the latest technologies and information sets there exists a window of opportunity during which those technologies and information sets are not yet “priced into the market”.

So are markets efficient?

Yes, maybe, and no. Yes, markets are probably efficient with respect to traditional structured data sources processed by traditional investment models. Maybe, markets are efficient with respect to traditional structured data sources processed by the latest forecasting models. No, we believe that markets are probably not efficient with respect to structured and unstructured data sources processed by the latest forecasting models.

The Philosophy of Computational Investing

Our investment philosophy at NMRQL Research is bold and simple: continuously do machine learning at an unprecedented scale on exponentially increasing sets of structured and unstructured data. In so doing we will constantly be adopting and using the latest forecasting models, information sets, and search technologies. If we can do this faster than our competitors, we believe that this approach will enable us to consistently beat the market because we will be operating in the window of opportunity before new technologies are adopted widely enough to be priced into the market.

Investing in Exponential Times

The characteristics of technology to improve exponentially has resulted in us living in what can best be described as exponential times:

“An analysis of the history of technology shows that technological change is exponential, contrary to the common-sense “intuitive linear” view. So we won’t experience 100 years of progress in the 21st century — it will be more like 20,000 years of progress (at today’s rate)” – Ray Kurzweil in 2001

Whilst we don’t agree with Kurzweil that this inevitably results in the technological singularity, we do agree that this exponential speedup is happening. What we believe at NMRQL Research is that investing in exponential times requires an entirely different mindset and a more pragmatic and less dogmatic approach to investing.

Three years in and we would say we are still in phase one of our quest to execute this vision. Ultimately the success of any investment manager lies not in their philosophy, but in their execution and the performance of their funds as measured by return and risk. On this basis, NMRQL Research looks forward to being tested in the financial arena.