AI Asset Management Report

How Can AI innovate asset management?

Innovation in asset management is difficult to be detected from outside.

Investors, who are clients of an asset manager, are interested in the brand and performance of asset managers, but they are relatively disinterested in the way how their assets are managed in practice. Because of the zero-sum characteristic of the asset management industry, asset managers intend to hide the hard-won innovation from competitors for as long as possible. Thus, innovation in asset management is difficult to be detected from outside.

Contrary to common conception, numerous innovations have been made quietly in the field of asset management, and the hunt for the next innovation is still in progress. , When we first look at what kinds of innovations took place before looking into the possibility of AI-driven innovation in asset management, we may be able to get a clue about whether AI technology is a viable innovation driver in the asset management area.

Daily Contrarian Trading Strategy: STRATEGY C

Let’s think of a  simple investment strategy called a contrarian strategy. (Let's call it ‘STRATEGY C’ for convenience.)

Before the opening of the market, find the stocks that rose (“winners”) and fell (“losers”) on the previous day.

Long losers, with a higher portfolio weight assigned to the stocks with a greater decrease in performance (i.e. buy more stocks that had fallen more) Short winners, with a bigger portfolio weight assigned to the higher outperforming stocks.

Assemble a market-neutral portfolio with the same amount of long and short. Rebalance this portfolio every day.

This seems to be a very simple strategy, but the return of this strategy is outstanding.

* Amir E. Khandani and Andrew W. Lo, 2007, "What Happened To The Quants In August 2007?"

The table above exhibits that this simple daily contrarian strategy delivered averaged 1.38% daily returns during 1995. (If the same strategy had been implemented only in the smallest market-cap group, it would have returned 3.57%, daily.) The portfolio would have increased by 30x when 1.38% daily return was annualized (3,000%!). The sharp ratio is 53.87, which means this portfolio is virtually risk-free.

STRATEGY C was first published in 1990, and it is known that hedge fund managers adopted it slightly earlier. Early Quant funds, such as D.E. Shaw & Co., and Renaissance Technologies, made a huge fortune with STRATEGY C and were able to level up their houses at the early stages of its history. Famous Quant Funds such as PDT Partners (which used to be Morgan Stanley's proprietary trading group), had also made a lot of profit with STRATEGY C.

The daily return of STRATEGY C had diminished to just 0.13% by 2007 from 1.38% in 1995, while many managers scrambled to deploy the same strategy. The daily volatility doubled to 0.72% from 0.40%. However, many quant funds are still actively using a revised version of this strategy (e.g. neutralizing the portfolio by industry to reduce risk, or to find a better universe or condition where the strategy works better).

The research paper, ‘What Happened to The Quants in August 2007(Khandani and Lo)', shows that during August 2007 many high-profile Quant funds (using a daily contrarian strategy) experienced unprecedented losses that were triggered by the rapid unwinding of one or more sizable quant portfolio, despite relatively little movement in the equity market. (The co-author of the paper Andrew Lo is a Chinese professor at MIT Sloan School who went to Harvard and wrote the first paper covering the contrarian strategy) James Simons of Renaissance Technologies has a history of excusing that “the funds’ sudden underperformance appears to be the result of rapid unwinding from other quant funds with similar strategy”

Two questions arise here.

Why such a simple and powerful strategy could only be discovered in 1990?

Why has the strategy worked well for such a long time, after it was first announced in 1990?

The answers to both questions are actually the same, and the answer is deeply related to innovation in asset management through AI technology.

From individual stock to portfolio: the beginning of Statistical Arbitrage

STRATEGY C was simple yet very powerful, eliminating the impact of the market (beta) by buying and selling the same amount of winner and losers every day. Many people had put a lot of effort to research how to make money, but it was not until the late 1980s that few people only noticed it and started making big money with this strategy. What was the reason?

In fact, to implement STRATEGY C, all you need is a daily price data for all stocks and a simple computer for programming. These two conditions were met even in the 1950s. However, the necessary conditions are different from sufficient conditions, and there is a big difference between what was available and what is achievable actually. In the 1980s, the probability that the one who has access to all the past stock price data and the one can test the strategy C with a computer programming is the same person was close to zero.

Back in the 1980s, star managers such as Peter Lynch and hostile M&A raiders (the background of 'Pretty Woman') were at their peak, and most investors were just focused on how to pick better stocks and obtain high-profile (or sometimes insider) information. (It is not really different from these days.) At the time, computer and price data were only used importantly at IT departments dealing with back-end systems. Individual stock analysis functions and chart functions were enough for most of the investment departments. The number of Bloomberg terminals broke through 5000 only in 1986, and the first MS Excel for Windows 2.05 (it was not for 3.1) was released in 1987. In other words,

people were focused on analyzing individual stocks, not portfolio strategies.

The necessary conditions for Strategy C were there, but the environment was not sufficient for anyone to deploy STRATEGY C easily.

Even after Andrew Lo published a paper on STRATEGY C in 1990, investors paid little interest, and STRATEGY C's daily return remained above 1% even in 1995. There were still only a handful of teams with the will and capacity to back-test and execute STRATEGY C on the entire stock universe, who moved away from the stock-picking approach. These teams that moved first and fast, while keeping the secrets, made a lot of money.

The first movers were computer scientists and mathematicians such as D.E. Shaw, James Simons, and Edward Thorp: Wall Street outsiders who were trying to make money with a scientific approach. They were the outliers who could deal with computer technology and finance at the same time, for the first time in history. As these teams discovered innovative strategies like STRATEGY C, the quiet innovation in asset management has begun. Not long after, this type of investment strategy started to be called statistical arbitrage, and most quant funds use a statistical arbitrage strategy nowadays.

Innovation is difficult to start, but once it starts, it progresses rapidly.

The Era of Quant Fund: The Birth of Quant Fund and a New Balance

As the first-generation Quant Hedge Funds continued to show outstanding returns, quant hedge funds became bigger and bigger. As a result, more than half of the highest-paid hedge fund managers, including James Simons of Renaissance Technologies, who ranked top with an annual salary of $1.8 billion in 2018, were filled with managers using computer algorithm trading.

Morgan Stanley’s proprietary trading group that is known as one of the first teams to start statistical arbitrage, Renaissance Technologies by James Simons who was a mathematics professor, D. E. Shaw & Co. by Shaw, a computer science professor (and once worked for Morgan Stanley's proprietary trading group) and Englander’s Millennium Management are 1st generation quant funds who started computer algorithm trading and statistical arbitrage trading in the mid-80s. And those who were from the first-generation company independently established second-generation quant funds.

WorldQuant founded by a game programmer from Millennium Management and PDT Partners led by Peter Muller, a mathematician from Morgan Stanley's proprietary trading group (with the introduction of Volker rule after the financial crisis, Morgan Stanley's prop trading group was spun off as PDT Partners), and Two Sigma founded by a silver medalist at International Mathematical Olympiad from D. E. Shaw & Co. are the second generation quant funds.

As the number of quant funds increased and their AUM increased, problems began to arise. In a market with zero-sum attributes, most strategies have had short lives with a sharp drop in returns as large-scale Quant funds quickly find similar strategies and roll out huge amounts of money at the same time. The return erosion from the increased competition outpaced the speed of quant funds coming up with new strategies. The name of the game has changed to race for faster research with more and more quant analysts on board than other competitors in a labor-intensive way. As a result, the number of employees hired by Quant funds has increased, and for the top Quant funds, average AUM per person has fallen to around $30 million.

Top Quant Fund's AUM per Person

The combined remuneration and overhead costs of each employee at Quant Fund exceed $500 k. If $30m is managed per person, the fund will have to charge at least 2% to keep the fund running smoothly. Thus, most Quant Hedge funds are operated based on a 2-20 structure, (charge 2% of fixed fee and 15-20% of excess return for performance fee). Cheap alpha no longer exists.

Automated Quant Research

As explained earlier, the high fee structure is inherent at Quant funds for its expensive approach of seeking alpha; hire more highly paid analysts. Elites from Ivy League at Quant Funds try to find an excess return strategy by organizing data, preprocessing it, and back-testing many ideas. If there is an idea that "the momentum strategy works well after company disclosures", they would try to find alpha through multiple versions of back-testing and forward-testing the idea. Numerous reiterations are required from different angles: which universe works better, which disclosure works better, which measure-based momentum strategy works better, how long it works better after disclosure, etc. However, those trials are to little or no avail most of the times.

IF

1) we can expedite the speed of research on excess return strategies

2) Furthermore, if it is possible to automatically generate the portfolio management strategy without employing an army of highly paid researchers.

THEN

It is possible to offer alpha at a lower cost to wider investors base than it does now. This is a sure way to advance and secure a superior market position in the rapidly growing Active Index market.

To see if such innovation is possible, it is necessary to define the problem more clearly.

The problem of automatic generation of investment strategies and AI

f(X,U) = P

X: Dataset    U: Investment Universe    P: Performance

Simply put, finding a portfolio management strategy is like finding a function f that is expected to perform well in the future for investment universe(U) and input data(X). For example, if you breakdown the S&P Index (“P”), the investment universe (“U”) is [US large-cap stocks], the input data (“X”) is [market cap], the function (“f”) is [invest at market cap ratio, with quarterly rebalancing].

To overly simplify what Quant funds do with the above formula, they try to find a desirable function (“f”) by experimenting with various candidates of “U” and “X”.

 For "X", the candidates were much simpler in the past such as the price data and the financial data of all individual stocks as well as macro data – e.g. interest rates, exchange rates, indices, and economic indicators. Many researchers still use only these three types of data as an X candidate. For example, the simple function “f” to buy the bottom 10% of the stock with the lowest PBR with annual rebalancing just requires stock price and net asset value as input data, “X”.

However, the simple function above, like STRATEGY C, no longer performs well. This is because too many investors are already aware of and using it. It is necessary to find an investment strategy “f” that is not easy to find, faster than others to enjoy steady excess returns in the zero-sum market.

More specifically,

- [Data differentiation]: When using data X that others do not see well as parameters.

- [Investment Universe Differentiation]: cases where the investment universe is dynamically defined.

- [Function differentiation]: If "f" is complex or has a nonlinear relationship.

when any of those above conditions are met, other researchers would not be able to detect your winning strategy for a considerable period.

1. Data Differentiation

Attempts to access proprietarily source and differentiate data source sounds fancy, but surprisingly few have been successful. This is primarily because no matter how differentiated (often private and unstructured) data is, data that is not related to the movement of the actual portfolio is useless. , There are not many private data  that actually contains the rich alpha source. The following examples explain it well.

A big quant fund, launched its own satellites to measure the size of the glacier, for the use in natural gas futures trading (which is sensitive to temperature and climate change). But it eventually abandoned the project due to the lack of fund performance. All satellites had been sold.

Attempts to use Walmart's parking lot image data from satellites for trading eventually failed.

A hedge fund, which traded the news by measuring the sentiment for each stocks with natural language processing, announced that it had to change the strategy because of poor performance.

Another hedge fund, founded by a professor led team that intended to use Twitter mention data for trading, was opened with great fanfare and the attention of the media but was quietly closed for lack of performance.

We cannot simply generalize the above cases, of course. Using good and effective data that are hidden will obviously give you an edge. However, looking at the unsuccessful attempts thus far, it seems clear that attempts to find a better strategy with public data can be more successful than attempts to gain an advantage from hidden data especially for finding alphas with high capacity.

The reasons are; 1) many of the unstructured data lag the stock price; 2) there is a high probability of overfitting due to insufficient data samples or difficulty in conducting back-tests over a long period of time, and 3) the alpha deposited in the data was actually not very big, presumably.

When you search through a vast amount of data (most of which is unstructured) like Google users' search data, you may feel these new loads of information might be of good use for stock trading. But the level of true alpha you can extract out of such new data is most of the time smaller than simple price data.

2. Function differentiation and Investment Universe Differentiation

If you can derive very complex patterns that others have not seen before from the same given data set, your chance of seeking alpha increases. The problem is that the human cognitive ability is not really designed (or configured) to recognize/understand nonlinear patterns.

A linear pattern like “if you invest in low PBR stocks, they will likely to rise in the future, or if you invest in underperforming stocks, you’ll get higher returns”, fits the human cognitive structure better. However, suppose the stock price follows a random formula below with a considerable probability, it is difficult to spot/identify such pattern (or formula), especially when there’s always some level of noise in the data in reality.

A random nonlinear formula

Even at a lower degree of complication for example, the phenomenon that the predictive power of PBR varies depending on the size of the company is a very simple nonlinear pattern that is a source of alpha works pretty well, is not easy to find. Let us remind ourselves that Joel Greenblatt's simple magic formula, which recorded a return of more than 40% per year for 20 years from 1985 to 2005, was not easily spotted. (Magical formulas, just for the like of STRATEGY C, could be discovered by Greenblatt only with the introduction of computer technology in the 1980s that availed easier back-testing)

In other words, to success in function differentiation, we need tools that can easily find nonlinear patterns. Just as we needed the organized data sets and computers to discover Strategy C back in the 1980s.

The same goes for investment universe differentiation. Qraft Technologies (“Qraft”)published the results of a study that momentum and value factor investing works well for US Large caps especially a month or two post the company filing/disclosures. For human to easily find the same result (and even more so when done automatically), an efficient tool is required to easily back-test what happens to a specific pattern for one to two months period post the public disclosure for individual stocks in the US. Of course, without an efficient tool, you can back-test the strategy via multiple layers of complex codings, theoretically.

However, there is a big difference between finding such a successful strategy from scratch and back-testing the already-discovered strategy. What you “could” find is different from what you have found. Just as people were not able to notice a simple STRATEGY C (although anyone “could” have done before) when data and a computer were not easily available. It is impossible to find a winning strategy like above, without an efficient tool that can handle a dynamic investment universe (in this case, the universe of US stocks that have just made the announcement).

Qraft was able to find this winning strategy for the dynamic investment universe because we have an efficient tool that can handle the investment universe as a “variable”, not a “constant”.

The well-designed Deep learning model works best for finding function “f” reflecting both of these nonlinear relationships and dynamic investment universes.

3. Dimensions

There are tremendous combinations of functions to back-test. There are thousands of data fields available for quant researchers. Considering the degree of freedom of an investment universe that could be examined, the number of functions that can combine each data field with various investment universe is infinite. It is similar to the Go game. The brute force method, which tests and rules out all the cases possible to discover a function (strategy), is impossible.

An experienced quant researcher, like a professional Go player, can narrow down the number of cases to be tested by intuitively figuring out the relationship between the stock performance and the data (to avoid painstaking brute force). This shortlisted thesis (or function) is again tested to derive a candidate for winning strategy. This increases the probability of finding an investment strategy. To automatically extract a good strategy without an experienced quant researcher, you need to come up with another way to narrow down the vast search space. AlphaGo solved this problem by applying several techniques, including deep learning technology, and outstripped human abilities.

Deep learning technology can solve the problem of finding an optimal function (investment strategy) in a huge search space as exhibited by AlphaGo case.

4. Overfitting

Overfitting must be handled to ensure the quality of the strategy. It is dangerous to fit the model using all the data available. The result of back-testing would be fine but the same result cannot be achieved in practice which is the out-of-sample test. Especially in the case of financial data, it is not easy to cope with overfitting because the length of the time series data is short and the characteristics of the market change frequently.

It takes a lot of time for human researchers to find a model that fits well with all the previous data sets available. Hence some human researchers create a model using the entire dataset (ignore overfitting) and try to reduce the risk of overfitting by assessing the reasoning (i.e. the rationality of the strategy).

If you build a system that automatically finds the function f by applying deep learning, you can make the system learn to use the data released before the specific time of prediction  (inference) only. This greatly reduces the probability of overfitting.

5. Rusty Strategies

Strategies built by Quant research methods are basically static strategies. That is, the new data flooding in every day are not reflected in the strategy. If the strategy deviates from the market and does not fit well, you have to either discard it or maintain it with new data. However, the investment strategy created with a DL model learns new data every day, and the weight produced by the neural network changes as per the new data set, thus the life of the investment strategy is much longer. (Even deep learning models are subject to new model engineering if and when the new type of data sets are introduced)

Automated Strategy Extraction by AI

The quant fund operation is largely composed of; data processing, strategy research, and order execution(trading).

In the data processing stage, complex financial data is processed to remove any bias for a better simulation. At the strategy research stage, numerous research analysts are employed to find the alpha source and the optimal portfolio, using pre-processed data from the previous stage. Sophisticated order execution (trading) is essential to minimize the market impact (and the transaction cost) when processing large scale trading orders.

Qraft Technologies has built an AI-based system for all of these 3 steps above.

source: Qraft Technologies, Inc.

The overall system structure is as follows.

source: Qraft Technologies, Inc.

The most important element in creating this automated system is how you can construct an environment where you can test the complicated models as easy as you can. Automation can be achieved only if this condition is met.

1. DATA PRE-PROCESSING SYSTEM: QRAFT KIRIN API

It is very difficult to use raw financial data from vendors such as S&P Global and Refinitiv (formerly Thomson Reuters). Elimination of survivorship bias (e.g. treatment of delisted stocks), look-ahead bias (e.g. correct treatment of revised financial statements) and accurate processing of corporate events such as M&As, rights offering take private and re-listing, take a tremendous amount of time to pre-process. Qraft Technologies' data processing systems automate the pre-processing tasks through parallel computations accelerated by GPU.

(https://youtu.be/bsxUPNTSDjA)

Qraft data processing system not just collects and stores data, but enables to test the investment universe from various angles with pre-processed data sets. For example, you can define the investment universe as "companies that have not passed two months since the latest public disclosures of patents," and test the pertinent investment strategy with just a few lines of python codes. (In other words, Qraft data processing system allows us to find an investment strategy automatically). This system is packaged as API and soon to be commercialized as a standalone business solution.

2. STRATEGY EXTRACTION SYSTEM: QRAFT ALPHA FACTORY

The core function of the AI research system is the automatic extraction of investment strategies. The huge amount of search universe multiplied as [investment universe (U) x degrees of freedom] * [data (X) x degrees of freedom] * [function form (f) x degrees of freedom] is much more massive than the search universe of a Go game. In such vast search universe, a well-engineered deep learning model has exhibited consistent results in narrowing the probable candidates and automatically back-/forward-testing the candidates to finally extract an investment strategy. The automatic extraction of investment strategies through deep learning is composed of two modules.

1) Factor Factory

Factor Factory is a system that automatically searches basic patterns with potential for excess returns by applying AutoML technology. Using one NVIDIA DGX server, Factor Factory can produce more than 10 patterns (factors) per day without human intervention. It consumes a lot of electricity, but you don't have to pay high payroll and there is no one leaving the job.

Example of factor and factor performance extracted from factor factory (Source: Qraft Technologies)

*Qraft Factor Factory rediscovered well-known factors in academia without any supervised learning. 
https://bit.ly/2VTxhQV

*For more information on Qraft Factor Factory, see the following Qraft newsletter link.
https://bit.ly/3oyO9si

*Our article on Medium also covers Qraft Factor Factory in more detail.
https://bit.ly/3lT9Cuf

2) Strategy Factory

Factors found through Factor factory are not independent sources, so it is more appropriate to construct nonlinear combination models as opposed to linear combinations. Strategy factory extracts the investment strategy with a nonlinear asset price model by nonlinearly combining factors extracted automatically from the factor factory. The more factors accumulated through Factor factory, the more sophisticated asset price models that strategy factory generates.

3. AI Execution System (AXE)

The history of the order execution system by computer algorithm is very long. However, most automated order execution systems use predefined rule-based algorithms such as VWAP, TWAP, and IS. JP Morgan Chase announced the world's first AI order execution system, LOXM, which applied deep learning reinforcement learning technology to tick data of individual stocks. Shortly after, Goldman Sachs announced that it has also developed and tested the AI order execution system.

Qraft Technologies also developed an AI order execution system that applies reinforcement learning technology to stock tick data. It is the globally first case of commercialization and provided to Shinhan Investment Corp. (JPMorgan and Goldman Sachs use it for internal use only and do not provide the systems to the outside.) Qraft Technologies was able to apply the latest AI-enhanced learning model developed during the relatively recently. Unlike JP Morgan's LOXM, which was created using the DQN model used in AlphaGo, Qraft Technologies has applied the latest agent-based reinforcement learning model to significantly increase the performance of order execution. (https://youtu.be/o3m6Ewjc7Lw)

DRL model applied to Qraft Technologies' AXE

In 2018, during the AXE Challenge sponsored by NVIDIA, Shinhan Bank, KOSCOM (total prize of US$100 K) the same large-cap portfolio (randomly generated by KOSCOM) was given to the securities dealer and AXE.  The challenge was to see who could buy the portfolio cheaper over a week. The challenge was broadcasted live on TV, AXE beat the human dealers from securities brokers by a large margin. 

Shinhan Financial Investment (South Korea's 2nd largest banking group) are using AXE since March 2020 for executing equity orders of the National Pension Service, showing outstanding results compared to VWAP.

AXE explores the optimal order execution strategy by learning patterns from tick data, including the price and transaction volume, as well as the history of transactions including order types. When AXE is applied, it is possible to improve the  return of active index funds by minimizing the transaction cost of mass orders of all financial products. It especially works more effectively with super-sized funds or small-mid cap funds that have a larger market impact.

Qraft Technologies recently announced that it is packaging AXE as a cloud solution with Microsoft. When the cloud solution is ready, all financial institutions will be able to apply AXE to their services through the API connection.

With AXE, a messenger platform will be able to replace MTS that has a complicated order process UI with a brokerage service based on natural language processing of "Buy $50K worth of Apple stocks at the best price over the next week”. When this new type UI/UX spreads with the help of AXE, the complicated MTS that requires various order types such as order prices, market/specified prices, etc., would lose its place.

NVIDIA, the world's largest GPU company, selected Qraft Technologies as one of the 30 AI startups globally as ‘Inception Premier’. Qraft Technologies is the first Korean AI company to be selected for NVIDIA's Inception Premier, and the only firm in the financial service AI category globally. (Other companies selected are unicorns or famous AI startups that had been acquired by Apple.) Most companies in the list consist of self-driving, vision-oriented AI companies that require a lot of GPU processing.

NVIDIA's pick of Qraft Technologies as the first premier member in the financial sector might be of NVIDIA’s business prospect that GPUs can be sold to more financial institutions when AXE penetrates to the global financial market. (For AXE, more the customers’ orders, the more real-time simultaneous processing power required by AXE for GPU-powered parallel learning and inference process)

Is this the beginning of another 1990?

Back in 1990, a few pioneers had found Strategy C with the advent of data and a computer.

In 2020, another group of pioneers could achieve a big innovation in the asset management industry, “automated alpha”, with the help of AI technologies.

AI-driven asset management model

AI-driven asset manager model only needs a team of data engineers for discovering and dealing with new data, and a team of AI engineers to build deep learning models that avail highly efficient strategy extraction. And the cost is limited to subscribe to data sources and tick data sources for order execution. No matter how big the number of funds to be developed and operated, there is no additional cost other than the server expansion. It is not a model of the far future, but a model that is being implemented right now.

The team that can introduce an efficient strategy extraction (Alpha factory) that human researchers could never achieve, would be the right candidate to innovate the investment management industry. A new pattern of investment strategy (e.g. dynamic investment universe, just like surprising AlphaGo’s next move that was beyond human perception) will be derived and the same team could make good active index funds and those funds will seize the significant portion of the asset management market. Perhaps that this might be the very beginning of innovation to unlock inexpensive alpha with AI technology in the asset management market that is growing towards $10 trillion.

Qraft Technologies is

A company using AI technology to innovate the inefficiency of the asset management industry. From data processing to alpha research and portfolio order execution, we aim to solve inefficiencies at each stage of the asset management industry with AI technology. With this, we can provide a high level of alpha at a lower cost.

We automate complex financial data pre-processing, accelerate it through parallel computing, and automatically search for the alpha factor through AutoML technology in a well-established simulation environment. Using the alpha factors found in this way, a deep learning-based Deep Asset Pricing Model is created through the Strategy Factory as per the fund universe defined for a fund concept. The final portfolio made through this model is efficiently executed with the order execution engine AXE based on deep reinforcement learning.

Qraft Webpage: http://www.qraftec.com

QRAFT

Subscribe Qraft Newsletter: https://www.qraftec.com/newsletter 

Newsletter — QRAFT

Disclaimer

The past performance may not be indicative of future results.

This material was prepared for informational purposes and cannot be used for the purpose of soliciting the sale of financial investment products such as funds.

This document contains the contents of the patent-pending or registered by Qraft Technologies, Inc.